Changelog
[2.32.0] - 2026-11-13
Section titled “[2.32.0] - 2026-11-13”2.32.0 (minor) — mods can tell whether another player is online, a player’s profile no longer goes blank in mods after a quick lookup, and outfit areas reset to default no longer read as black. Adds SocialSnapshot.Presence / IsOnline / Sections and the SocialPresence / SocialSections types (Abstractions 2.32.0); additive, no plugin rebuild.
- Mods can now tell whether another player is online, when they went offline, and the status they have set.
- A player’s ability score, class, gear and outfit no longer disappear from mods after the game briefly looks that player up again (for a nameplate or portrait).
- An outfit area the player reset to its default colour no longer shows up as black in mods.
Developer notes
Section titled “Developer notes”-
SocialSnapshot.Presence(SocialPresence?, init-only), decoded fromSocialData.basic_data:OfflineTimeMs(field 9, Unix ms, 0 while online)OnlineTimeMs(field 16)PersonalStates(field 8, packed or unpacked;ChatStatusTableids, never null)ReceivedAtMs(stamped by the cache)IsOnline(OfflineTimeMs == 0, the game’s own rule)
The shortcut
SocialSnapshot.IsOnline(bool?) is null whenPresenceis unknown. (#165) -
SocialSnapshot.Sections(SocialSectionsflags) records whichSocialDatasections a snapshot holds.Nonemeans not reported (hand-built snapshot). (#165) -
Cache merge:
SocialDataCachenow merges per section instead of letting the last reply win.- A section the reply carries replaces the cached one, even when it is present but empty.
- A section the reply leaves out is kept.
Presence,LocationandPartycarry forward.Sections == Nonekeeps the old replace-everything behaviour.
This fixes thin replies (basic-only, nameplate) wiping
FightPoint/ProfessionId/Gear/Fashion/Identity. (#165) -
Dye hue decoding (
Stellar.Wire): a dye’sxis hue degrees only for a Board colour group (ColorGroupTable.Type == 3). For other group types it is a swatch id looked up in the group’sHuerows. A miss,x == 0orColorGroupId == 0decodes to hue 0, matching the client (fashion_data.lua:327-351).- New
DyeHuePalette+DyeHueMapper, with additiveRead(span, DyeHueMapper?)overloads onAttrFashionDataReader/SocialDataReader. FashionDyePaletteResolvercaches per fashion id and per group id, and reflects only on the main thread. A miss during an I/O-thread decode falls back to degrees.- In release_3.7 every dyeable piece uses a Board group, so this only changes the 4 group-0 pieces and guards future builds. (#164)
- New
-
{-1,-1,-1}reset sentinel: a reset area is now dropped fromFashionEntry.Dyes/DyeAreas(the piece’s default colour) instead of decoding to black. (#164) -
Compatibility. All additions are init-only properties and new types, with no positional-record change, so existing plugins keep working. A plugin that reads
Presence/IsOnline/Sectionsneeds framework ≥ 2.32.0.
[2.31.0] - 2026-11-13
Section titled “[2.31.0] - 2026-11-13”2.31.0 (minor) — the wardrobe 3D preview can show the weapon and each outfit’s own hidden pieces, and lists scroll with the mouse wheel while you drag a row. Adds IWardrobePreviewOptions (IWardrobePreview.ShowWeapon and a Show overload with a hidden-pieces map) (Abstractions 2.31.0); additive, no plugin rebuild.
- Wardrobe mods can now show your weapon in the 3D outfit preview, including the weapon skin saved with an outfit.
- The 3D outfit preview can now show an outfit exactly as saved, with its own hidden pieces, even when you are hiding a piece (like a hat) right now.
Changed
Section titled “Changed”- While you drag a row to reorder a list, the mouse wheel now scrolls the list, so you can move a row past what’s on screen.
- Sliders that can’t be used right now look clearly switched off: a sunken track and a faded knob, at the same size.
Developer notes
Section titled “Developer notes”IWardrobePreviewOptions— a new base ofIWardrobePreview(which stays at the STELLAR0005 8-member cap; nothing moved, so compiled callers are unaffected; only the framework-internalWardrobePreviewServiceimplements it). Reached asservices.WardrobePreview:bool ShowWeapon { get; set; }— defaultfalse(today’s weaponless preview). Live, no model rebuild. Global to the preview service, so set it before eachShow. With an outfit that carries a weapon skin (WardrobeRegions.WeaponSkinPreview) the model wears exactly that skin (the existing social-data injection + CMount de-dupe); without one it wears the player’s current weapon.void Show(EntityId self, IReadOnlyDictionary<int,int> outfit, IReadOnlyDictionary<int, IReadOnlyDictionary<int,float[]>>? dyes, IReadOnlyDictionary<int,bool>? hidden)—hidden= region →true(hidden) over the 14WardrobeRegions.Allregions; a missing region is shown, other keys are ignored. The framework writes the model’s ownEWearSetting(all 14 regions explicit) after dressing, so the player’s current global hide setting no longer leaks into the preview.null= the existingShowexactly. The[WardrobePreview] weapon …log line gainshide=ok(rb=same)(orrb=<read-back>/hide=err:…) when a map was passed — check it on the first in-game test, the live client’s Lua can differ from the extracted one.
- Reorder + wheel (#162): during an active
ListElement.OnReorder/VirtualListElement.OnReorderdrag the ticker appliesInput.mouseScrollDelta.yto the list content with the ScrollRect’s own wheel step (scrollSensitivity), clamped, gated on the pointer being over the list viewport andFrontWindowBlocks. The press-drag scroll lock stays; a press that hasn’t become a drag yet still ignores the wheel. No allocation; no work on frames without a wheel delta. - Slider disabled look (#160): when
SliderElement.Enabledreturnsfalsethe slider paints the recessed rule-22 look — track [email protected] (translucent, blends with the surface), fillMenuMuted@0.30 (value still readable), knobMenuMuted@0.70, no ring, no size change. It flips live withEnabled, repainting only on a state change (no per-frame graphic writes) and on theme change. The stockSelectabledisabled tint is set to white so it no longer multiplies on top. The enabled look is unchanged; plugins that faded or swapped a disabled slider themselves can drop that workaround on ≥ 2.31.0. - Compatibility. A plugin that sets
ShowWeaponor calls the 4-argumentShowneeds framework ≥ 2.31.0.
[2.30.0] - 2026-11-12
Section titled “[2.30.0] - 2026-11-12”2.30.0 (minor) — a redesigned Settings window and launcher, the Sakura theme, more movable game HUD parts, and less CPU in raids and crowded places. Adds IThemeStatusColors, ThemePreset.Sakura, TextElement.MaxLines, IFrameworkDeveloperMode, ICombatBuffEffectLookup, WindowSpec.Prebuild/Topmost/FixedHeight (Abstractions 2.30.0); additive, no plugin rebuild.
- A new built-in theme, Sakura: rose-purple panels with a pink accent.
- Your own themes can now change the text colour, the muted text colour, and the green, red and amber used for “on”, “remove” and “careful” states.
- Mods can now show text over a set number of lines, ending in “…” when it is longer.
- Drag your favorite plugins to put them in the order you like, in the Stellar menu’s Favorites tab or on the compact bar. Both use the same order.
- Five more parts of the game HUD can be moved and hidden in the layout editor: Status Icons, Buff Timers, Item Bar, Action Buttons, and the location and line bar under the minimap (the line-recycle notice the game sometimes shows there moves and hides with it).
- The layout editor shows a dashed placeholder for game HUD parts that aren’t on screen right now (for example the boss HP bar outside a boss fight, or the party panel while solo), so you can place or hide them at any time.
- A “Hide others” switch next to the layout editor’s SHOW filter: while you edit one plugin or only the game UI, everything else is hidden so it doesn’t get in the way. It is never saved and everything comes back when you leave edit mode.
Changed
Section titled “Changed”- The Stellar Settings window has a new look: tabs down the left side, a short status line at the top of each tab, and a message at the bottom confirming every change you make. It keeps one size, so it no longer grows or shrinks between tabs.
- Every setting has a “?” button that explains what it does.
- Plugins show their own icons, and a plugin that failed to start is marked red with a Retry button.
- Themes are picked from cards, with a live preview window that follows your theme, opacity and classic button style. Custom theme colours open in their own window.
- Hotkeys are grouped by plugin, show “Not set” when unassigned, and ask before resetting everything.
- Game UI shows which parts of the game HUD are moved or hidden, and each can be reset on its own.
- Performance lets you pick each plugin’s rate from a list.
- The launcher shows bigger plugin icons, five plugins per row, and every tile is the same size whatever its name.
- New, sharper icons throughout Settings and the launcher, drawn for each size they appear at.
- Buttons you can’t press now look sunken, so they are easy to tell apart from buttons you can.
- The window opacity slider sits with your theme and is remembered per theme; font and UI size moved to the General tab.
- The compact bar shows plugin names in normal white text, and its Settings tile matches the other tiles.
- Dropdowns in Settings use the same colours as the other buttons.
- Switching between Settings tabs, and opening or closing Settings and the launcher, is much smoother.
- Mods use less CPU in raids and other crowded places, so busy fights run smoother.
Removed
Section titled “Removed”- The “Reset selected to default” and “Reset all in this slot” buttons on the Layout tab, which did nothing. The layout editor’s own toolbar resets windows.
- Your own themes based on a theme that this version doesn’t know are now kept, instead of disappearing the next time settings are saved.
- You can now have up to 12 of your own themes. Before, a 13th theme was saved but never shown, so you couldn’t select, rename or delete it.
- A custom theme can no longer be named like a built-in theme (for example “Default”), so it can’t be mistaken for it. Existing themes keep their names.
- Scrolling lists in mod windows is smoother.
- Text in mod windows stays sharp when you just click a window or a list without moving it, and moving, resizing or drag-scrolling slowly no longer flickers between sharp and soft.
- The Stellar Settings window and the Stellar launcher open instantly the first time, instead of freezing the game for a moment.
- Plugins can now have their big windows prepared while you log in, so they open without a freeze the first time.
- Scrolling with the mouse wheel now stops a list that is still gliding after a drag.
- Fewer small stutters from mod windows while you stand idle, for example when a new chat message arrives.
- Scrolling a mod window list with the mouse wheel no longer stutters on every step.
- Dragging or scrolling the Stellar menu, Settings or chat no longer makes other mod windows stutter.
- Thai text in mod windows now stacks tone marks above vowels correctly (for example ปลั๊กอิน), and all Thai text uses the same font.
- Renaming one of your own themes keeps its colours, and Duplicate copies the selected theme’s colours instead of starting from its base theme.
- Settings hints now show the keys you actually set in Hotkeys, not the default keys, and say so when a key is not set.
- Game HUD parts you moved no longer jump back to their default spot for a moment after a scene change, a menu or a cutscene; the chat window in particular stays where you put it.
- Game HUD parts you hid no longer flash back on when the game shows them again.
- The skill bar and the player HP bar can now be hidden too. Skills and HP keep working normally while they’re hidden.
- A hidden chat window no longer opens an invisible input box when you press Enter.
- The skill bar’s edit outline now covers the whole bar, including the dodge buttons.
Developer notes
Section titled “Developer notes”- Native game-UI wrapper ownership (
NativeUiService/PandaHudAdapter, framework-internal). Moved/hidden game HUD nodes are now owned through a frameworkstellar_wrap_<id>RectTransform (outer wrap for in-prefab nodes; inner wrap — wrapper as the element’s only child, adopting late children — for view roots chat / target frame and the class gauge). Position = wrapper offset, hide = wrapper scale 0 (neverSetActiveon game nodes; chat also blocks input via a CanvasGroup on the wrapper). Removed: the 1 Hz re-assert + idempotent guard +Resources.FindObjectsOfTypeAllresolve scan; replaced by a 10 Hz alloc-free intact check and scene-enter re-wrap. Saved rects migrate once to offset keys (.ox/.oy, marker.mv). AllowlistRectChildspecs gain a-exclude token;*unions walk through plain Transform holders. Layout editor: placeholders, “Hide others” (IWindowEditHide, transient, never persisted). No Abstractions/SDK change. - Developer mode readable by plugins (
IFrameworkDeveloperMode, a new base ofIFramework):bool DeveloperMode { get; }andevent Action<bool> DeveloperModeChanged, reached asservices.Framework.DeveloperMode. Read-only — plugins cannot switch it. Meant for showing a plugin’s debug tools / verbose logs only while the user has developer mode on. The event fires on the main thread on every flip (Settings, or a hand edit ofstellar.framework.config.json→developer.developerMode), not at subscribe time.PerPluginFrameworkholds ONE subscription on the sharedFrameworkServiceand fans out to the plugin’s handlers;Unregister(plugin disable / unload / failed ctor) drops it and clears the plugin’s handlers. Each plugin’s handlers are isolated (a throw is logged with the plugin guid; other handlers and plugins still run).IFrameworkstays at the STELLAR0005 8-member cap (nothing moved, so compiled callers are unaffected; only implementors — the framework-internalFrameworkService/PerPluginFramework— change). Needs framework ≥ 2.30.0. - Hotkey hints follow the live binding (
HotkeyHintText, internal). The Settings texts that named a framework hotkey hard-coded its default: About footer (Shift+Home), Layout footer + the “Edit mode” panel caption (Shift+— never even the default, which is Alt+E), the Developer tab's perf label + its help (Shift+End). They are now{0}templates (settings.foot.about,settings.foot.layout,developer.perf.label,help.developer.perf) filled with the action's currentKeyBinding.ToString()(exactly the Hotkeys key pill), with*.unboundlines for a cleared key; the Layout caption shows the binding orhotkeys.notSet. Cached per template, dropped onBindingChanged, a language switch, or a new declaration (the directory's action-list snapshot changes — a later default can silently take a framework chord). Render-driven only. The unusedlayout.editEnter/layout.editExit` keys (dead since the hub redesign) are removed. All six languages. - Buff effect ids (
ICombatBuffEffectLookup.TryGetBuffEffectIds(EntityId, int buffUuid, out IReadOnlyList<int>)), a new base ofICombatLookup(which stays at the STELLAR0005 8-member cap; nothing moved, so compiled callers are unaffected; only implementors — framework-internalCombatServiceand the testStubCombat— change). The combat wire readers now keep what they used to drop: PlayEffecteffect_ids (logic type 0),BuffEventCustomize(1002)values,BuffEffect.HostUuid,BuffEffectSync.Uuid,BuffInfo.HostUuid(4) andBuffInfo.LogicEffect(13, incl. inside an AddBuff(18) payload), plus the appearEntity.buff_effect(field 8, ids only — its buff upserts stay ignored). Rules = resonance-logs-cn decoder.rs (AddBuff payload ids → outer PlayEffect ids when an AddBuff/BuffChange entry is present → Customize values; Remove / a RemoveLayer that ends the buff drops them), with the ExperimentMcBuffEffectPatchquirks kept (absent PlayEffect RawData = id 0; whole-buffereffect_idprobe before the 1-/2-byte customize fallback). Stored in aCombatServiceside table keyed by (uuid >> 16, BuffUuid), own lock, 4096-entry valve, cleared withClearAllBuffs(scene change / logout); NOT a field onActiveBuff(positional record = binary break). New internal sink memberICombatBuffSink.ApplyBuffEffectIds. Allocation: anint[](and the update list) only when an effect actually carries ids; decoding uses stack buffers. Lets Mechanic Callouts drop its three game hooks (parseBuff/OnAddBuff/OnBuffEffectSyncprefixes). Needs framework ≥ 2.30.0. - Later 2.30.0 additions (this branch).
ConditionalElement.Prebuild/PrebuildName(time-sliced background build, 3 ms/frame),ConditionalElement.KeepAlive(parked branches keep their nested canvas, bindings gated; instant re-show) andConditionalElement.ContentRevision(background refresh of a parked branch when its content changes). Settings hub: background warm off-screen until settled, soft park on hide (canvas off + off-screen instead of rootSetActive(false)), hidden visible-tab refresh. Launcher re-warms in the background when its content signature changes while hidden (e.g. world entry). Per-size pixel-hinted framework icons (Resources/icons/<12|16|24|32|64|128>/,LauncherIcons.Get(name, px); exact size draws 1:1, else next larger with mips). Thai: TMPmark/mkmkpositioning + one-batch Thai glyph seeding + un-wrapped GPOS lookups in the shipped Thai fonts; legacyTextshowing Thai renders through a TMP twin in the shipped face. Hidden Developer tab (developer.developerModeconfig key or 7 clicks on the About version; Shift+End perf overlay is developer-only). Theme store:IColorOverrideStore.RenameTheme/CopyTheme/RemoveTheme,CustomThemeStore.Duplicate;ThemeOverrideSnapshotfor Save/Discard. All additive; measured in game at 2560×1440 (tab switch +10–15 ms, open/close one ~2× frame). IThemeStatusColors(Success,Danger,Caution) is a new facet onIThemeColors(IThemeMenuColorsstays at 5 members). Only the framework implements these interfaces, so this is additive. Dark presets return exactly the old hard-coded house values#5FD38D/#FF4D4F/#FFCC73; Light returns#1E8A4C/#D13438/#A86A00. Status indices 18..20 are appended toThemePresets.Tables.- Editable framework slots:
Theme.MenuText,Theme.MenuMuted,Theme.Success,Theme.Danger,Theme.Caution(labelstheme.color.*in all six languages), so custom-theme overrides feed theMenuText/MenuMuted/status reads. Older frameworks keep but ignore these override entries (ColorOverrideStorenever prunes unknown slots). ThemePreset.Sakura = 4(appended). Menu#2E2433/#E8DCEF/#9B8AA5/ accent#C8508F, gradient#2E2433 → #291F2E → #30242F, default window opacity 0.90. A plugin that registers per-preset colours without a Sakura entry gets its Default colours.ColorRegistryService.Register(key, label, defaultAll)now fills every enum value.CustomThemeStore: an unparsable (or undefined numeric) base preset loads as Default and its original string is written back on save.MaxCustomThemes = 12+CanCreate; a 13thCreatethrowsArgumentException; existing configs with more load in full.TextElement.MaxLines(int,init, default 0 = unlimited). The framework breaks the lines itself (WordWrapCap, shared by the TMP and plainTextpaths): only at spaces, after hyphens and between CJK characters; a word wider than the line is cut with “…” on its own line, never split across lines; the last line ends in “…” when more text follows. The engines’ own wrap is OFF for such text (never TMP’s Ellipsis mode). Re-fitted on a string, font-size or width change.TextElement.Height(float,init, default 0 = the text’s own height): a hard-fixed box height (min = preferred), lines centred in it; pair withMaxLines.ScrollElementnow registers its content with the pixelPerfect-while-scrolling logic (previously only virtual lists did), so scrolling no longer re-snaps and rebatches the whole canvas every frame. Internal, no API change.- pixelPerfect vs drag/resize:
WindowInteractionTicker.TickWindowMotionlatches real motion of the dragged window (anchoredPosition) or resized target (sizeDelta+ position) frame to frame (> 0.5 px), seeded on the first held frame and cleared on release, so a press on a drag handle / resize grip without movement never flips pixelPerfect. Held-gesture rule (HeldMotionRule, unit-tested), shared by window move, window resize and list drag-scroll: the gesture’s FIRST real move (> 0.5 px) turns its canvas’s pixelPerfect OFF, and every frame of the rest of the hold counts as motion, pauses included, until the mouse button is released; then it comes back ON 0.15 s after the release (lists: after the glide ends; a wheel tick during the glide stillStopMovement()s it). Press-and-release without movement = no flip. Replaces the earlier “re-crisp 0.15 s after holding still mid-drag”, which flipped at every pause: one slow drag-scroll of the launcher grid logged 17 flips oncanvas=settings.hub(≈2 ms write + 15 ms rebuild each, 71728ac). Per-view hold state isScrollView.Engaged(cleared withPressed); the window gesture’s is_motionEngaged(cleared when the drag/resize ends). Per-canvas routing and flip logging unchanged. Internal, no API change. - Internal house-style kit
Stellar.Infrastructure.UI.Style(StellarStyle,StellarHelpPopover,StellarFeedbackLine) for the upcoming Settings hub and launcher redesign. No public API. - Disabled pills are recessed (user-approved option A, replaces the transparent + 1 px muted ring look): a disabled
PillButton/NeutralButton/PrimaryButton/SmallPillkeeps its size and radius and draws a SUNKEN fill,StellarStyle.DisabledFill(surface)= black @28 % over the surface it sits on (the panel for 28 px pills, the list tile for 22 px row pills), faded by window opacity like every fill, with the label at theme Text @32 % alpha, raised in 4 % steps until it reaches 2:1 on that fill (DisabledLabel(surface)). No outline ring (DisabledOutlineremoved). No inner top shadow: a Bubble has one child and no top-only stroke, so it would need an extra overlay layer per pill. ImageElement.Live(bool,init): re-pullPngevery refresh and swap the texture when the returned array reference changes (the launcher TileElement IconBinding path, decode deduped per window).ImageElement.Tint(Func<ColorRgba?>?,init): poll-diffedRawImage.colorfor flat white glyph PNGs (null = untinted). Both default off = unchanged.- Launcher redesign (house style, approved prototype):
LauncherViewsplit into.Full/.Tabs/.Tile/.Compactpartials over the internalStellarStylekit; list logic inLauncherMenuModel(tabs, Find above 16 plugins, Favorites ghost). Newlauncherconfig keytab(Favorites/All; missing = Favorites); pins unchanged (favorites = pins). - Settings hub redesign (phase 2).
SettingsHub(+.Rail) builds the fixed 800 × 612 window (1 px height strut +VerticalFillflex fillers; tabs lazily built perConditional),SettingsKithands every panel the kit, help popover, footer feedback line andStellarRowFlash; each panel exposesDescribe(SettingsKit)+Status(SettingsKit)(the one status builder). “Edit colours” is a child window (stellar.settings.colours, opened centred on the hub and fronted); help popover windowstellar.settings.help. Recon walk moved to About → “Save HUD paths”. InternalBootstrapPlugin.OpenSettingsAt(string tabKey)(keysplugins/layout/themes/hotkeys/gameui/performance/about) for the launcher.IPluginDisplayNames.IconOf(default member) feeds plugin launcher icons to the list. Kit additions:SmallPill(22 px),SmallCloseButton(24 px),FilledTile,VerticalFill,TopColumn,MinHeight,Border(); fixed-width pills now setMaxWidth. All new strings in the six framework languages (additive keys only; superseded keys are left in place). - Settings hub pre-warm. The hub (
stellar.settings.ugui) is mounted once in the background (Wiring.Settings.Prewarm.cs: a transient_hubPrewarmingflag AND-ed into itsShouldRenderandHubShown; Visible for 3 ticks, then parked). It starts once the window canvas scale is ready, outside the loading screen, pastGamePhase.Startup, and once the client language is latched (ClientLanguageLatch.IsKnown, or at CharSelect). The cold first open measured a 316 ms frame (mount 235 ms,build:firstApply230 ms) at 2560×1440. The mount’s first apply also builds the starting tab’sConditional. It is never fronted (ZFront stays 0); real opens still go throughShowSettingsHub(un-park = fresh ZSeq, thenBringToFront). Skipped whenSTELLAR_AUTO_OPENnames the hub. - Launcher pre-warm (same recipe, now shared). The one-shot logic moved into
WindowPrewarm(Host;Warming/Settled/Tick(control, ready)/EndForRealOpen), pumped byWiring.Prewarm.cs(wasWiring.Settings.Prewarm.cs) for both windows, ONE at a time: the launcher (settings.hub) first, the hub only after the launcher has settled, so the two builds never share a frame. The launcher’s cold first open measured a 193 ms frame (build:firstApply149 ms,mount:settings.hub149 ms). ItsShouldRenderANDs!Warminginto the loading gate;ToggleLaunchertreats a warming launcher as closed (opens it, never hides it), ends the warm-up beforeSetVisible(true), then fronts as before. It builds the persisted Full/compact mode and Favorites/All tab (Conditionals onLauncherRegistry, visibility-independent). Same gates as the hub (canvas scale ready, not loading, past Startup, language known or CharSelect); skipped when already shown or named inSTELLAR_AUTO_OPEN. - Wheel / press stops list inertia.
WindowInteractionTicker.TickScrollInput(.ScrollViews.cs) callsStopMovement()on a coasting StellarScrollRect(ScrollElement / virtual / variable list / dropdown viewports) on any wheel delta or mouse press over it. uGUI only zeroed the velocity on BeginDrag. The internalRegisterScrollViewhook now also passes theScrollRect. It respectsFrontWindowBlocksand an open dropdown popup. The game’s own ScrollRects are untouched (no Harmony patch). - pixelPerfect flips only on USER scrolling. Before,
TickScrollActivitycounted any contentanchoredPositionmove over 0.5 px as scrolling, including programmatic ones (a chat re-pin to the bottom, a list growing or shrinking, rebinds). Every such move flipped the root canvaspixelPerfectoff and back on, which is 2 whole-canvas rebuilds. The fw 69a41c4 log showed 88 idle[Hitch]lines in about 10 minutes (ui:canvasRebuild30–50 ms,ui:tickeraround 8 ms in the same frame). Now motion counts only while the user moves that view continuously (IsUserScrolling): the left button held since a press on it (content or scrollbar drag), or fling velocity (ScrollRect.velocity, which only a user drag creates; our views areClamped, so no elastic velocity). The mouse wheel never counts: a tick moves the content once, one ordinary re-snap with pixelPerfect ON, while counting it cost OFF + ON per tick (measuredpixelPerfect OFF reason=scroll:settings.hub write=5.3ms canvasRebuild=25.0ms, back ON 0.3 s later ≈ 60 ms of stutter per tick). A wheel during a glide still callsStopMovement()(zeroes the velocity), so pixelPerfect comes back ON after the normal 0.15 s settle. The per-viewWheelTime/ 0.3 s wheel window is gone. The drag/resize motion latch is unchanged. Scroll views are now one entry each (ScrollView: content, motion and input state;AddScrollView), replacing the parallelScrollContents/_scrollPrevYlists. - Flip instrumentation (
[UiPerf], armed with HITCH / diagnostics). Each flip counts asui:ppFlipon the[Hitch]line and asppFlips=Nin the 10 s summary. A throttled line (at most one every 2 s) logs it on the next frame:[UiPerf] pixelPerfect OFF reason=scroll:<windowId> write=0.4ms canvasRebuild=41.5ms frame=63.2ms (+2 flips not logged). The reason isdrag:/resize:/scroll:<windowId>/rest, andwriteis thepixelPerfectsetter itself. - Window stacking is no longer sticky (
WindowStacking, unit-tested). Before,BringToFrontset a separate monotonicZFronttier that outranked category and mount order forever, so once the launcher and Settings hub fronted themselves on every open (this branch), plugin windows that never front themselves opened UNDER them. The renderer token now carries one global activation value (ZAct), and the sort is (bottom → top): popups (DismissOnOutsideClick) >Topmost>ZOrder> layer > activation. The layer is HUD below Tools/Debug; a HUD window that wasBringToFront’d is lifted into the Tools/Debug layer until it is next hidden (ParkclearsZLifted), which keeps the documented “HUD above Tools” use without stickiness. Activation happens on: first mount; a REAL show from hidden, meaningSetVisible/SetVisiblePersistrising edge (Entry.ActivateOnShowmakes the un-park call the newIWindowOrder.Activate);BringToFront(any category); andSetTopmost(true). Re-activation is NOT triggered by visibility restores (ApplySavedRect(applyVisibility)on mount / slot switch,ResetRect), which use the new internalEntry.SetVisibleRestore, so they never reshuffle z. Un-parking no longer refreshes order by itself. Tools and Debug now share one layer (Debug no longer always sits above Tools). A HUD shown from hidden still goes on top of the HUD layer only, as its old freshZSeqdid. ExplicitBringToFrontis unchanged for callers: still needed for an already-visible window, harmless otherwise.BringToFrontPending(front before the first mount) still applies at mount.FrontWindowBlocksand every ticker hit-test read sibling indices, so input follows the new order.OverrideSortingOrderwindows and the dropdown popup root canvas are separate canvases above all of this, unchanged. The pre-warm never fronts; a real open mid-warm-up fronts explicitly. WindowSpec.Topmost(bool,init) andIWindowTopmost(bool IsTopmost { get; },void SetTopmost(bool)), implemented by every framework window handle. It is a separate interface becauseIWindowControlis at the STELLAR0005 8-member cap. Call it through theWindowTopmostExtensions.SetTopmost(this IWindowControl, bool)/IsTopmost(this IWindowControl)extensions, which cast and are a no-op/false on the inert duplicate-id handle. A runtime toggle before mount, or across a scene-change remount, is re-applied at mount. Intended for small pinned overlays/alerts. Needs framework ≥ 2.30.0.- Per-canvas pixelPerfect. A window with its own nested canvas (
WindowSpec.IsolateCanvas, orOverrideSortingOrder) now getsCanvas.overridePixelPerfect = true+pixelPerfect = trueat attach (WindowRenderer.AttachNestedCanvas), so at rest it looks exactly as before. The ticker (WindowInteractionTicker.PixelPerfect.cs) keeps one switch per canvas (PpTarget: the root, plus one per isolated window, registered viaAddLocalPixelPerfect, pruned with the window). Each motion source is routed to the innermost isolated window it lives under (Transform.IsChildOf): a window drag/resize target, or a user-scrolled/gliding scroll view’s content. Only THAT canvas is flipped OFF and settled back ON; the shared root canvas, and every other window on it, is not rebuilt. Motion in non-isolated windows flips the root as before, and each canvas settles independently (same 0.15 s settle and user-motion rules). Flip logging now names the canvas:[UiPerf] pixelPerfect OFF canvas=root|<windowId> reason=… write=…ms canvasRebuild=…ms(UiPerfStats.PixelPerfectFlipgained acanvasargument). The rebuild probe times Unity’s whole canvas update, so nested canvases are covered.IsolateCanvasis now ON for the launcher (settings.hub), the Settings hub (stellar.settings.ugui) and the “Edit colours” child (stellar.settings.colours). That is a good reason for any big scrolling plugin window to setIsolateCanvastoo. WithoverrideSortingfalse, a nested canvas still stacks by sibling index (stacking /FrontWindowBlocksunchanged) and has its ownGraphicRaycaster.FireOnClickWithRectandPositionDropdownnow readrootCanvas.scaleFactor, since the nearest canvas can be a window’s nested one. - Launcher favorites drag-to-reorder. The Full menu’s grid is now one 5-column
ListElement(94 × 102 cells,Gap6: the same geometry as the old pooled rows) withOnReorder;CanReorder=LauncherMenuModel.CanReorderShown(Favorites tab, no Find text, not a ghost). Reordering is OFF while Find narrows the grid (a filtered grid hides tiles, so the drop slot in the full order is ambiguous). The compact bar’s favorites are oneListElement(1 column, orHorizontalfor the horizontal bar) withOnReorder; Settings sits outside it, so it is never dragged nor a drop target. Order persistence: newlauncherconfig keyfavorite_order(string[]of titles) next to the unchangedpinnedset, an ordering only (LauncherFavoriteOrder.Sort: saved ids first, then pins it lacks in registration order, so a config without the key looks exactly as before). A new favorite appends at the end (which writes the order the first time); un-favorite removes it; a pin migration renames it in place; favorites whose tile is hidden right now keep their slot (LauncherFavoriteOrder.Apply); a ghost keeps its slot across a reorder and re-starring puts it back there. Feedback: the dropped tile flashes; the footer showslauncher.feedback.moved, and the Favorites idle hint becomeslauncher.hint.idle_favorites(all six languages). The framework’s press lock on the ScrollRect means a reorder drag never scrolls the grid and never counts as list motion for pixelPerfect. 13 tests (LauncherFavoriteOrderTests).-ListElementgained three init properties (Abstractions 2.30.0, additive):CanReorder(Func<int,bool>?, polled on the press: false = the press stays an ordinary click, still a drop target; a throwing predicate locks the item),Gap(float?, null = the old 2-unit spacing, both axes for a grid),Horizontal(bool: one HorizontalLayoutGroup row sized to its active items; a fixed-column GridLayoutGroup would reserve every pooled column; reorder insertion is left/right). StellarStyle.ConfirmPanelis exactlyConfirmSpec.Widthwide (new init property = the panel’s OUTER width where it sits). The bubble getsMaxWidth = Widthand both texts a fixedWidthof the content (Width − 20), so they wrap; the pills are a plain left-aligned row at a fixed label width (min(96, (content − 8)/2 − 24)) instead ofPillPair’s weighted cells. Before, a confirm nested in a column panel hugged its one-line sentence and pushed the column wider (Themes delete confirm: text past the edge, pills ~165 px each into the right column, the caption hint clipped to “2 of”, the card grid’s scrollbar poking out). All four confirms set it: Themes delete and Layout remove =ColumnWidth − 20, Hotkeys / Game UI reset all =MainWidth − 20.-CustomThemeStore.IsBuiltInName:Create/Renamethrow for aThemePresetname (any case, trimmed); the Settings rename field also refuses the preset names as shown in the current language, withthemes.nameError.builtin(all six languages). Loading never drops an existing theme with such a name.- Colour picker child window (
stellar.settings.colourpicker, option B).ThemesPanel.Picker.csregisters it next to Edit colours (GlassMenu, no title bar, Draggable,IsolateCanvas, fixed 260 wide =ThemeEditorBody.PickerContentWidth236 + 2 × 12); content =ThemeEditorBody.DescribePicker: title strip (slot name, ellipsized, + round ✕) over the unchanged sharedColorPickerElement(live edit, persist on release). A row’s Edit pill reports its screen rect (StellarStyle.WithClickRect→SelectableElement.OnClickWithRect);ThemeEditorBody.EditRowrecordsEditAnchor+ bumpsEditSeq.TickPicker(fromTickColoursWindow): every fresh open DOCKS it at Edit colours’ right edge + 8 (left when there is no room), top at the clicked row (pill centre − 13, screen px × Canvas/Screen), clamped on screen, re-asserted for 4 ticks (first-mount DefaultRect); docked, it follows Edit colours each tick; a rect that differs from the last one set = user drag → undocked until it closes; Edit on another row while open re-targets in place (docked → realigned) +BringToFront(already visible). Closes via the row’s Close, its ✕, Done / Edit colours closing, the hub closing; its position is never persisted (colours.picker.titlefor the layout lists, all six languages). Edit colours is fixed height now: “Other colours” scrolls in a fixed 6-row area (166) and the open “+ Add colour override” section replaces it in the same height (its list 132). WindowSpec.FixedHeight(bool,init, default false): a GlassMenu window keepsDefaultRect.Heightinstead of content-fitting its height (verticalContentSizeFitteroff; ignored whenResizable). A body taller than that is squeezed toward its min height (list viewports have min 0) rather than growing the window. The Settings hub uses it (800 × 612), so no tab can resize it; the Performance tab used to (its per-plugin list reserved 7 rows, now 5, and the tab’s preferred height ran past the 478 px tab budget).WindowSpec.Prebuild(bool,init, default false) +WindowSpec.PrebuildSignature(Func<int>?,init): build a hidden plugin window ahead of its first open. Framework-driven (plugin ticks don’t run at the login screen): the element tree is built at login / character select WITHOUT any value pull (only build-time callbacks run: buttonLabel/Icon, sliderGet,TitleProvider), then one off-screen fill (every value lambda, Conditional, list) per world login — back to back behind the world-entry loading screen, else in play outside combat 3 s apart. Also fills a window that is Visible but held hidden byShouldRenderand never drawn yet (targetlens.bufflist / statinspector.mini-hudfirstshow74 / 45 ms); its takeover isShouldRenderturning true. The window stays hidden (IsShownfalse for a hidden one, nothing persisted, no clicks/hover); a real open mid-fill takes over; a changedPrebuildSignaturerefills while hidden. Raid log 11-10 first opens it targets: dungeonhub.confirm.raid 418 ms, dungeonhub.hub 204 ms, chattools.main 160 ms, combatmeter.main 121 ms. Register the window at start-up, not lazily. Needs framework ≥ 2.30.0 to set.- Launcher re-warm behind the loading screen. The 240 ms
mount:settings.hubin the 11-10 raid log was the launcher’s own title-screen warm (settings.hub = launcher), not an open. What reached play was its world-entry re-warm (1 → 22 tiles,apply:settings.hub76.9 ms, frame 111.7 ms): it waited for the loading screen to clear. The framework warms (launcher, Settings hub) may now run behind a loading screen onceIsWorldActiveis true under it. DropdownElement.FillColorFn/HoverFillColorFn/LabelColorFn(Func<ColorRgba>?,init, default null = the built-in look): Pill-style trigger colours, polled every refresh (final colours — apply window opacity yourself). Null hover = the built-in 8 % text wash over the custom rest fill; a custom label also tints the ▾ (diffed per refresh). Classic and the option list are unaffected.SettingsKit.Dropdownpasses the house Neutral fill / hover / label, so every Settings dropdown (Language, per-plugin Rate / Self-rate, theme preview) matches the Hotkeys key pill. Needs framework ≥ 2.30.0 to set.- Compatibility. A plugin that reads the status colours, selects Sakura or sets
MaxLines/TextElement.Heightneeds framework ≥ 2.30.0. SettingImageElement.Live/Tint,WindowSpec.Topmostor callingSetTopmost/IsTopmostalso needs ≥ 2.30.0.
[2.29.1] - 2026-11-09
Section titled “[2.29.1] - 2026-11-09”2.29.1 (patch) — PR #151. Mod buttons in the Esc menu’s side bar stack under the Stellar button with their own icons, and several mod-window drag and slider problems are fixed. No API change, no plugin rebuild.
- Buttons that mods add to the Esc menu’s side bar now line up one below another under the Stellar button, instead of sitting on top of it.
- Mod buttons in the Esc menu’s side bar now show their own icon in its original colours, like the game’s own buttons; only the Stellar button keeps its glowing star.
- Large round slider knobs now look perfectly round, instead of like rounded squares.
- Dragging a slider in a mod window that has no title bar now changes the slider, instead of moving the whole window.
- Dragging a mod window by its title bar no longer sometimes does nothing when a closed mod window used to sit in the same spot.
- A slider scrolled out of view in a mod list no longer stops you dragging the window around that spot.
- Click-through mod overlays, such as a raid alert banner, no longer stop you dragging or hovering the mod windows underneath them.
Developer notes
Section titled “Developer notes”- Clip-aware scrollbar/slider exclusion.
PointerOnScrollbarignored RectMask2D/Mask clipping, so a slider registered viaRegisterScrollbar(e048f00) that was scrolled out of a ScrollRect viewport still claimed presses past the viewport edge and suppressed window drags there. It now goes throughPointerHitRule.AcceptsPresswithOnInflatedBar && !OutsideClip(same rule as hover/HitTest/HitSlot); the 6 px horizontal inflation is unchanged. - Passive windows no longer cover in
FrontWindowBlocks. A visibleWindowSpec.Passive(click-through) window in front counted as covering, so e.g. RaidManager’s wideraidmanager.mech.alertbanner blocked drag + hover on windows behind it.WindowRenderer.Mountnow records Passive roots in the ticker’sPassiveRoots;FrontWindowBlocksskips them unlessLayoutEditGate.IsEditing(PointerHitRule.FrontWindowCovers) — in edit mode they stay selectable/draggable and still cover. Internal only, no public API change. - Hidden windows no longer steal title-bar drags.
HitWindowHandlewas the one ticker press hit-test without anactiveInHierarchyguard. A hidden window is parked (rootSetActive(false), drag registrations kept) or ShouldRender-suppressed, and its handle still hit-tests at its last geometry; a parked Tools window sorts above HUD and has no active window in front, soFrontWindowBlockspassed and the press dragged the INVISIBLE window (the visible one stayed put, the hidden one moved silently). Now routed throughPointerHitRule.AcceptsPresslike the other press hit-tests. Pre-existing since window parking (62fd71b, v2.16.0), not a regression of this branch. - Slider drag-exclusion. On a title-bar-less
Draggablewindow the whole frame is the window-drag handle, andWindowInteractionTickeronly exempted scrollbars, so a press on aSliderElementstarted a window drag.BuildSlidernow registers its rect through the sameRegisterScrollbarexclusion hook (FrontWindowBlocks-checked). - Slider: the SquareHandle knob is now a true circle at any HandleSize. The knob reused the
Capsule9-slice (fixed 7-texel corners), so anything over 14 px drew as a rounded square. The SquareHandle path now setsImage.pixelsPerUnitMultiplier = 14 / HandleSize(CapsuleGeometry.CircleMultiplier) so each corner isHandleSize/2. The default stretched knob is unchanged. - Rail stacking.
PlaceBelowRailmeasured only native template siblings, so everyMainMenuRailMenuButtonSpecgot the same y.UGuiInjectionServicenow passes an internalUGuiInjectContext(RailSlot, Branded)toIUGuiCanvasAdapter.Inject; slot N sits N+1 steps below the native top group. Slots come from the registration list (branded first, then registration order; removed registrations free their slot), not from which siblings exist at inject time, so re-injects can’t swap or overlap. A Warning logs if a slot reaches the bottom Support/Unstuck/Exit group. - Branded vs plain. The framework’s Stellar rail button registers through the internal
UGuiInjectionService.RegisterBranded; it alone gets the accent tint + pulsingstellar-glowhalo. Plugin buttons (publicINativeUiHost.Register) drawIconPnguntinted with no halo, and keep the hover scale-up. No public API change (MenuButtonSpecuntouched);LoginSidebarunchanged.
[2.29.0] - 2026-11-06
Section titled “[2.29.0] - 2026-11-06”2.29.0 (minor) — PRs #148, #149. The mod framework now supports fully rounded pill dropdowns, drag-to-reorder in long lists, and more player party data for mods to read. It also fixes dropdown hover highlighting the window behind and centres pill dropdown text. Adds DropdownElement.CornerRadius, VirtualListElement.OnReorder and SocialSnapshot.Party (Abstractions 2.29.0); additive, no plugin rebuild.
- The mod framework now supports fully rounded ends on pill-style dropdowns, so mods can show capsule-shaped dropdown menus.
- The mod framework now supports drag-to-reorder in long lists, so mods can let you rearrange items by dragging them, and the list scrolls on its own when you drag near the top or bottom.
- The mod framework now provides more of a player’s party data, so mods can read that player’s party info.
- The mod framework no longer highlights the window behind an open dropdown when you hover one of its options.
- The mod framework now centres the text on pill-style dropdowns vertically.
Developer notes
Section titled “Developer notes”- Release scope. PRs #148, #149. The Added items show up only once a plugin uses them. The first users are Maestro (dropdowns, reorder) and EntityInspector’s Party tab. The hover fix applies to every dropdown, Classic and Pill, so ChatTools and MultiTool pick it up as soon as players update. The caption fix applies to every Pill dropdown.
DropdownElement.CornerRadius(#148) (float,init, default0, Pill only).0keeps the current 3 px chip corners. A value above 0 sets the trigger’s corner radius. It is capped at height/2 − 1 (13 for the 28 px trigger) because the 9-slice bubble sprite’s minimum size is 2r+2, so r = 14 drew a 30 px trigger. Any value ≥ 13 gives an exact 28 px capsule. Classic ignores it, and the option list keeps its radius-8 card and radius-5 rows.BubbleElement.CornerRadiusdocs now note r = S/2 − 1 for a round bubble.- Pill caption (#148). The bold-13 caption sits 1 px below the centred text box so the capital letters line up with the pill’s centre. The ▾ chevron stays centred. The pill is still 28 px tall.
- Open-dropdown hover (#148). The popup sits on its own root canvas, but the ticker’s hover test only knew window z-order, so cells under the pointer in the window behind also hovered. While a dropdown is open, only cells in the popup layer hover now (
HoverBlocked). Popup rows skipFrontWindowBlocks, because sibling indices from two different canvases can’t be compared. VirtualListElement.OnReorder(#148) (Action<int,int>?, defaultnull, uniform rows only, ignored whenVariableHeightis set). It works likeListElement.OnReorder(remove-then-insert item indices, never fires on a no-op, Esc or a drop outside the list cancels, 6 px drag threshold) and uses the same ticker state machine. The drag follows the ITEM index across recycled rows (VirtualListBinding.First, pureVirtualReorderMath). Edge auto-scroll speeds up linearly to 8 rows/s within one row of the top or bottom edge.ListElementbehaves the same as before: its drop math moved into the sharedVirtualReorderMath.RemoveInsertTo.SocialSnapshot.Party(#149) (new init-only property; additive, binary-compatible). DecodesSocialData.team_data(field 12,CharTeam) intoSocialParty { TeamId, LeaderId, TeamTargetId, TeamNum, MemberType, CharIds, Members }andSocialPartyMember { CharId, Name, Level, ProfessionId, TalentId, FightPoint, OnlineStatus, SceneId, EnterTime }. NewStellar.Wire/SocialTeamReaderhandles packed and unpackedchar_idsand falls back to the map key for a member’s char id.SocialIdentity.PartySizeis unchanged.SocialDataCachekeeps the last non-nullPartywhen a thin nameplate/avatar reply follows, the same rule asLocation. A[TeamDiag]line is logged only withSTELLAR_DIAGNOSTICS=1. Member order is random on every reply (proto map), so consumers must sort.MemberType0 = 5-player party, 1 = 20-player. A carried-forward party has no timestamp and may be stale.- Compatibility. A plugin that uses any of the three new members needs framework ≥ 2.29.0.
- Validation. #148: the build passes against the game interop and the CI stubs, and all tests pass except the 42 known Windows-only ReShade tests, including 28 new
VirtualReorderMathTestsand 11BubbleGeometryTests. The corner radius, hover fix, caption alignment and virtual-list drag-reorder were all tested in-game. #149: 24 new or extended social tests pass. In-game on 2026-11-04, the server sent the full roster for a player in a party you are not in: 5/5 and 9/20 parties showed every member’s name, level, class, online status and scene. A solo player comes back empty.
[2.28.1] - 2026-11-03
Section titled “[2.28.1] - 2026-11-03”2.28.1 (patch) — PR #146. The framework’s server clock now follows the real game server instead of your PC’s clock, so mods that time things against the server stay correct on a PC whose clock is wrong. No API change, no plugin rebuild.
- Mods that rely on server time (band sync, countdowns, cooldown timers) now stay accurate even when your PC’s clock is set wrong.
Developer notes
Section titled “Developer notes”- Release scope. PR #146.
ServerNowMswas the PC clock. World notify 43SyncServerTimeis{client_milliseconds = 1 (our own PC time echoed back), server_milliseconds = 2};PandaCombatStubProbe.OnServerTimeanchored on field 1. NewStellar.Wire/SyncServerTimeReader.TryReadparses the body in any field order and returns field 2, falling back to field 1 only when field 2 is missing or ≤ 0.ICombatSnapshotdocs now state the source and thatCombatEvent.TimestampMsis stamped from the PC clock (UtcNow), so the two differ by the PC’s clock error.- Effect. On a synced PC
ServerNowMsmoves by about one ping (~20–30 ms). On a PC with a wrong clock it now matches the game’sZ.ServerTime; consumers comparing it to PC-stamped event timestamps (CombatMeter settle/history/replay cap) shift by that error — optional plugin cleanup later. - Validation (in-game, 2026-11-03). PC clock deliberately 1.7 s slow: Maestro Link’s logged
Z.ServerTime − ServerNowMswent from+1742ms(2.28.0) to-3ms, and Link plays in sync. 6 newSyncServerTimeReaderTests.
[2.28.0] - 2026-10-29
Section titled “[2.28.0] - 2026-10-29”2.28.0 (minor) — PRs #143, #144. Mod dropdowns can use the rounded pill look from the Stellar Design Language, and a mod can give one of its windows its own background opacity instead of the global Window Opacity setting. Adds DropdownStyle / DropdownElement.Style and WindowSpec.FrameOpacity (Abstractions 2.28.0); additive, no plugin rebuild.
- Mod dropdown menus can now use a rounded “pill” look that matches the newer Stellar windows. Dropdowns that don’t opt in look the same as before.
- A mod can now let you set the background see-through level for one of its windows, separate from the global Window Opacity setting. Text and buttons stay solid.
Developer notes
Section titled “Developer notes”- Release scope. PRs #143, #144. Neither change shows up for players until a plugin uses it: the first users are MultiTool (Band Mute “Mode” dropdown) and ChatTools (a “Custom window opacity” slider), and both are released separately.
- Pill dropdown (#143). New
public enum DropdownStyle { Classic = 0, Pill = 1 }and theDropdownElement.Styleinitproperty, defaultClassic. It is not a primary-ctor parameter, so the record stays source and binary compatible (same pattern asMaxVisibleRows). NewWindowBuilder.Dropdown.Pill.cs:- Trigger: a 28 px rounded pill filled
Over([email protected], Over([email protected], MenuBackground)), with a bold 13 caption whose contrast is checked against the fill. Hover adds[email protected].Widthis honoured (0 = hug). Colours are read live from the theme, so a theme switch repaints it. - Chevron: an 8×5 anti-aliased triangle image. The bold TMP face has no “▾” glyph, so a text chevron gets a zero preferred size and never draws.
- Popup: a rounded panel with a 1 px edge and 4 px padding, and 26 px rows. The selected row gets a
[email protected]wash.MaxVisibleRowsscrolling still works. Positioning, the blocker, dismiss on Esc or an outside click, and the popup canvas are shared with Classic. - Refactor:
BuildDropdownScrollnow takes a(float H, float Gap) rowtuple to stay under the STELLAR0003 five-parameter cap. Classic passes its old values.
- Trigger: a 28 px rounded pill filled
- Per-window frame opacity (#144). New
WindowSpec.FrameOpacity, aninitproperty of typeFunc<float?>?, for GlassMenu windows only (Borderless windows already haveBackgroundOpacity).- Values: the delegate returns an override, or
nullto follow the globalIChromeStyle.WindowOpacity, so a plugin can switch between the two live. A null delegate keeps today’s behaviour. - Implementation:
ConfigureGlassRootgets its provider fromFrameOpacityOf(spec), which returnsClamp(own() ?? global, 0.3, 1)(the same 0.3 floor as the global slider). That feeds the existingFrameOpacityBinding. Only the frame Image’s alpha changes, andForceOpaque/ChromeKillare still honoured. If the delegate throws, the global value is used.
- Values: the delegate returns an override, or
- Validation (in-game, 2026-10-29).
- #143, on a local 2.28.0 package with the Default theme: MultiTool’s Band Mute “Mode” dropdown in Pill style showed a centred caption and chevron in a 140×28 pill, hover worked, the open list showed the selected-row wash, and picking a mode applied it. ChatTools’ “Translate to” dropdown kept the Classic look.
- #144, on a local 2.29.0 package with #144 rebased on #143, global opacity at 100%: ChatTools’ chat window frame followed the slider live (85% → 30% → 100%) while its text and input stayed solid. Other GlassMenu windows, including ChatTools’ Settings window, stayed at the global value, and returning
nullreverted the window. - Tests and gates: #143’s build had 0 errors, and its tests ran 3035 passed / 42 failed (all 42 are the known Windows-only ReShade tests). Neither PR ran
check-standards.shor #144’s unit tests locally; CI covers both.
[2.27.2] - 2026-10-27
Section titled “[2.27.2] - 2026-10-27”2.27.2 (patch) — PR #141. The Korean 한/영 key and live IME composition (Korean, Japanese) now work in Stellar text boxes, and the F1/F2/F3 quick-use item bar no longer disappears while you type in one. No API change, no plugin rebuild.
- The Korean 한/영 key now switches input language in a mod’s text box, the same as in the game’s own chat.
- Korean and Japanese text now shows while you are still typing it in a mod’s text box, instead of only after it is finished.
- The quick-use item bar (F1, F2, F3) no longer disappears while you type in a mod’s text box. The keys still type text and do not use items.
Developer notes
Section titled “Developer notes”- Release scope. PR #141.
- IME forced off every frame.
ZMouseManager.Update→updateMouseLockcallsPanda.ZInput.DefaultMousePointer.HideMousePoint()every frame unless a game view registered a show-mouse source; its body setsCursor.lockState = Locked,Cursor.visible = falseandInput.imeCompositionMode = Off. The setter icall is inlined, so it can’t be patched, and re-assertingOnevery frame (a01d827, reverted in 4ac9b6e) defeats the IME. NewImeKeepOnPatchis a Harmony prefix onHideMousePoint: while a Stellar field has focus (AnyFieldFocused, the same signalKeyboardInputGateuses) it re-applies the cursor lock + hide and skips the original, so IME Off is never written; otherwise the original runs. Installed fromApplyLifecyclePatches, unpatched by the framework Harmony host.AddShowMouseViewwas rejected: it switchesInputModeto UI and shows the cursor. - Composing text drew nothing. TMP draws one font per string, and
TmpRichInputFieldpicked it from the committed text only, while TMP appends the composition string to the label, not totext.TickComposition()now re-picks the font from committed text + composition while focused; the re-pick after a commit waits one tick so it doesn’t blink, and a cancelled composition falls back to the Latin font. - Quick-use item bar. Regression from 2.27.1 (a2483bc): the narrowed “UI” prefix rule stopped sparing
QuickUseItem(previously spared by accident via the bareuisubstring), and the game hides that HUD while its input mask is ignored.QuickUseItemis now in aMaskExcludeHudexact-name set; the keys stay blocked by the Rewired keyboard disable. - Validation (in-game, 2026-10-27). ChatTools composer, 한/영 simulated as
VK_HANGUL: before, 한/영 +gksgaveggks; after, it gives한, drawn live as ㅎ → 하 → 한; backspace-cancel leavesg; Japanesenihonnshowsにほんwhile composing and Space converts to日本; game chat unaffected. Item bar stays visible while focused, and F2 while focused used no item. Not checked: whether Alt still frees the cursor while a field is focused (screenshots don’t capture the cursor; Alt goes through the unpatched Show path).
[2.27.1] - 2026-10-27
Section titled “[2.27.1] - 2026-10-27”2.27.1 (patch) — PR #139. Auto Combo no longer stops while a Stellar text field is focused; ellipsized labels in virtual lists no longer flash a wrong cut while scrolling. No API change, no plugin rebuild.
- Auto Combo keeps fighting while you type in a mod’s text box. Before, it stopped casting until you left the box.
Developer notes
Section titled “Developer notes”- Release scope. PR #139. The ellipsis fix is only visible through the unreleased ChatTools whisper list, so it is deliberately left out of the player notes.
- Auto Combo while typing.
KeyboardInputGatemasked almost everyEInputMaskviaZIgnoreMgr.SetInputIgnore(EUIView) while a field was focused. The auto-battle AI cast path (ClientClickSlot→ZBattleUtils.ClickSlot→ZSkillInputMgr.onPress(isAIPress:true)) honours those masks. New exact-name exclusion set:NormalAttack, Skill, SkillWithWhiteList, DoByFuncId, Resonance1, Resonance2, Move, AutoMove, Walk, Run, LockTarget. Keyboard gameplay stays blocked by the Rewired keyboard disable andPlayerInputController.IgnoreKeyboard.QuickUseItemwas wrongly excluded by the bare"ui"substring rule and is now masked. Removed the deadcurFocusInputTextNamelookup (not aPlayerInputControllerfield). - Ellipsis on rebound rows.
TmpStyledText.MeasureAndFitwrote the new string (layout only marked dirty), thenFitToRectread the stalerect.widthleft by the slot’s previous string. The ~10 Hz re-apply corrected it about 100 ms later. Now, when the new string is wider than the rect,RelayoutForNewStringrunsLayoutRebuilder.ForceRebuildLayoutImmediateon the group picked byLayoutRootFor(climbs at most 4 levels while the width is content-driven) before fitting. Strings that fit cost nothing extra. The first attempt (0360b92) was a no-op in-game:transform.parent is RectTransformis usually false on IL2CPP, and845cdf8usesTryCast<RectTransform>(). Known leftover: one stale cut for about 10 frames when a tab first opens. - Validation. In-game: with the ChatTools composer focused and Auto Combo ON it kept casting (
IsInputIgnore(Skill)= false,Jump= true), and W/2 typed into the field without moving the character. The whisper list had no wrong cuts in a ~125 fps capture or in the owner’s scroll test.
[2.27.0] - 2026-10-26
Section titled “[2.27.0] - 2026-10-26”2.27.0 (minor) — PRs #135–#137. An opt-in setting keeps the mouse cursor free after you Alt + click a Stellar window. Clicks on Stellar windows no longer reach the game’s press-checks underneath. Also includes groundwork for the unreleased ChatTools redesign: the rich input caret, virtual-list row culling and a scroll anchor. Adds VariableListMath.TopVisibleIndex / FindRowById (Abstractions 2.27.0); additive, no plugin rebuild.
- New option in Settings → Layout → Cursor: “Keep cursor after clicking a Stellar window”. When it’s on, Alt + click a mod window and the cursor stays free after you let go of Alt, so you can keep using the window without the camera turning. Click the game world or press Esc to go back to the camera. That click or Esc only releases the cursor: it won’t attack, select a player or open the game menu. Off by default.
Developer notes
Section titled “Developer notes”- Release scope. PRs #135–#137. Fixes that are only visible through the unreleased ChatTools redesign (the press-check leak into the hidden chat input, the rich-input caret, row culling, the scroll anchor) are deliberately left out of the player notes. They become player-visible when that ChatTools ships.
- Rich input caret (#135, #136).
TmpRichInputFieldsetscaretWidth = 2. TMP’s default of 1 scales to about 0.05 × line height, which rasterized to nothing at the 13 px chat font. Separately, TMP’s own PageUp/PageDown scroll moved the caret rect and our text pin never re-synced it;AssignPositioningIfNeedednow does. A focus request restarts the blink (SetCaretVisible+ForceLabelUpdate). - Virtual-list row culling (#135). New
WindowBuilder.VariableList.Cull.cs, plus the same cull in uniform-heightWindowBuilder.VirtualList.cs. Latched rows fully outside the viewport areSetActive(false). The margin is max(2 estimated rows, 160 px) plus scroll velocity × 0.25 s. Rows keep their latched height and slot, so showing one again needs no re-measure. The pass runs only when the scroll position, viewport height or moving state changes;InvalidateHeightsforces it. Measured in-game at 2560×1440 with ChatTools open: Stellar renderers 835 → 427,clipCull0.60 → 0.27 ms/frame, mean frame time ≈ −0.2 ms. - Scroll anchor on trim/rebuild (#136). With
RowIdset, the top-visible row is tracked by identity and put back at the same screen offset after the consumer trims or rebuilds the list. New pure helpers inVariableListMath:TopVisibleIndex(the top row withoutFirstIndex’s pool clamp) andFindRowById(bounded outward search, at most 2·radius+1rowIdcalls). Both are unit-tested.RowId == nullbehaves as before. The natural >500-message ring trim has not been reproduced in-game yet. - Sticky cursor (#136).
StickyCursorService(Application) +IStickyCursorBackend→StickyCursorBackend(Infrastructure). It holds the game’s own show-mouse state (Z.UIMgr:AddShowMouseView). It arms on Alt + click on a non-HUD Stellar window and releases on a world click, Esc, world exit, setting off, or no Stellar window open. World-click swallow: a Harmony prefix skipsPanda.ZInput.ZTouchManager.simulateTouch(WorldPointerGate), and the Rewired mouse controller is disabled while armed (RewiredMouseGate) so the basic attack doesn’t fire. Esc is swallowed throughHotkeyKeyBlockPatch. Persisted asui.layout/sticky_cursor, default off, strings in all 6 locales. Tests:StickyCursorServiceTests. - Game press-check shield (#137).
Panda.ZUi.ZUIPointPressCheck(press_checkon about 90 Lua views) polls presses inUpdateand callsisPointerOverGameObject(Vector2). That method tests containment against everyEventSystem.RaycastAllhit, not just the topmost, so a Stellar canvas on top didn’t stop it. Observed symptom:main_chat_pcwas activated under ChatTools, and later hotkeys were typed into it.GamePressCheckShieldprefixes that method: when the topmost hit at the press point is a Stellar canvas, it returnsfalse. It uses the same test asIInputShield.IsPointerOverGameUi, so passive click-through windows stay click-through. It runs on press frames only, falls back to the original on error and is removed on teardown. It logs[PressShield] blocked game press-check under <windowId>(rate-limited). Side effect: a Stellar click now counts as an outside click for game views that close throughpress_check. Not shown in the player notes, because in-game the name card already closed on an outside click before this change.
[2.26.0] - 2026-10-25
Section titled “[2.26.0] - 2026-10-25”2.26.0 (minor) — first stable release since 2.23.1; rolls up the never-published 2.24.0–2.25.5 development versions. Lets mods show the game’s chat emoji inline in text and input boxes, rounded speech bubbles, smooth long scrolling lists, right-click menus and drag-to-reorder on picker cells, a “new messages” jump bar, per-person whisper conversations, the friends and private-chat lists, the game’s system notices and the chat send cooldown; injected name-card buttons get native dividers and fire on release only; config saves survive a transient file lock; fixes a multi-second login freeze. Adds API for plugins (BubbleElement, VirtualListElement variable-height mode, IPlayerProfile, IFriends, ChatMessage.PeerId/TargetId/SystemLabel/SystemType/SystemRefId, IChat.TryGetSendCooldown/SendResult, TextElement/InputElement.RichEmoji, InputElement.FocusRequest, InputElement.Placeholder/Locked, VirtualListElement.OnAtBottomChanged/ScrollToBottomRequest/RowRevision, SelectableElement.OnRightClick/Bare/OnClickWithRect, ListElement.OnReorder, RowElement.ConstrainsWindowMinWidth/OnHover, TextElement.Ellipsize/CenterGlyphs, BubbleElement.OutlineColorFn, IWindowHost.PrewarmText, HitchProbe.Count; Abstractions 2.26.0); additive, no plugin rebuild.
- Mods can show rounded, coloured speech-bubble cards that fit snugly around their text, including see-through colours, to match the game’s own look.
- Long scrolling lists of differently-sized rows now stay smooth however much history there is. Lists can start at the newest entry and stay pinned to the bottom unless you’ve scrolled up.
- Mod windows can now show the game’s own emoji as pictures inside text, and inside a text box while you type.
- Emoji added from a mod’s picker now go where your text cursor is, instead of always at the end of the text.
- Mod windows can show a “new messages” bar while you’re scrolled up; clicking it jumps back to the newest.
- Mods can now group whispers into one conversation per person, and show your friends list and private-chat list with unread counts.
- Mods can now show the game’s own system notices, such as “Obtained Guild Fund ×40”.
- Mods can now show how long until you can post again in World or Current chat, and tell you when a message was sent too quickly.
- A mod hotkey can now work only while that mod’s window is open. The rest of the time the key goes to the game.
Changed
Section titled “Changed”- Grids of picker items (such as a sticker or emoji picker) can open a menu on right-click, and you can drag items into your own order.
- Menus a mod opens from a button (such as a channel picker) can appear right next to that button.
- Buttons that mods add to a player’s name card, like Inspect, now have the same divider lines as the game’s own buttons.
- Buttons that mods add to a player’s name card no longer go off when you press on them to drag-scroll the card. They only respond to a normal click, like the game’s buttons.
- Mod settings are no longer sometimes lost when another program, such as antivirus, briefly locks the settings file during a save.
- Older chat messages that load when you log in now show under the correct channel in mods.
- Fixed a freeze of up to several seconds right after logging in while your recent chat history loads into a mod.
- Pop-up menus in mods no longer flash somewhere else and close the first time you open them after starting the game.
- Chat messages, whispers and who you whisper are no longer written into the mod log file.
- After switching characters, whispers the new character received while offline are no longer missing from mods.
- Replying to a whisper you sent now goes to the person you were talking to, not back to yourself.
- Mod windows can no longer be made too narrow for their own tabs or buttons, which cut them off in some languages.
- Hovering or clicking just outside a scrolling list no longer highlights or activates a half-hidden row.
Developer notes
Section titled “Developer notes”- Release scope. First stable since 2.23.1 = PRs #130–#133. The 2.24.0–2.25.5 sections were development versions that never reached the CDN; their developer notes are kept below, prefixed with their old version. Player-facing fixes to features that never shipped (2.25.3–2.25.5 list jitter/gap/overlap, rich-input centring,
StableSlotsredraw, notice ring,MarkPrivateReadbadges, reorder ghost) are deliberately left out of the player notes. - Name-card action dividers (#130). The game draws the thin divider as a static child of each native action item’s prefab (
ui/prefabs/idcard/idcard_item_tpl_pc); injectedIProfileCardActionsbuttons are built from scratch, so they had none.PandaProfileCardActionInjector.Divider.csfinds the divider on a native sibling (name contains line/div/sep, else a Graphic ≤ 6 px wide) andObject.Instantiates it onto eachstellar_action_*button asstellar_divider, stripped to visuals withraycastTarget = false. Missing divider → buttons build exactly as before (never throws). One-shot item-hierarchy diagnostics behindSTELLAR_DIAGNOSTICS=1. Validated in-game (release_3.7). - Name-card actions fire on release (#132). Injected name-card buttons fired on mouse DOWN, so pressing one to drag-scroll the action bar triggered it. New
PandaProfileCardActionInjector.Click.cs: still manual + level-based (GetMouseButton, the idcard canvas doesn’t dispatch EventSystem clicks to injected children), but press records button/position/rowanchoredPosition; the press is cancelled if the pointer moves >max(EventSystem.pixelDragThreshold, 10)px, the row scrolls > 2 units, or the button dies; release fires only inside the SAME button (drag checks re-run at release). Pending state resets on card close / re-inject. Cancelled presses logged underSTELLAR_DIAGNOSTICS=1. Validated in-game (release_3.7). - Config write retry (#131).
FileConfigStore.AsyncWrite.csdropped a write onIOException(seen in-game: an external transient lock, most likely Defender scanning the just-written file, lost a plugin’s “Clear all”). The writer thread now retries with backoff 50/100/200/400/800 ms (~1.55 s), abandoning the attempt if a NEWER save for the same path is queued (coalescing preserved); Info on a successful retry, one Warning (save failed after N attempt(s)) only when all fail.UnauthorizedAccessExceptionetc. unchanged (no retry). Shutdown drains share a 300 ms retry budget over ≤ 3 passes. The hash/stat ledger still records only after a successful write. Tests:FileConfigStoreRetryTests. - Whisper peer identity.
ChatMsgInfo.target_id(field 2, a charId) is now decoded (WireProtocol.TryParseChatMsgInfo5-arg overload;ChitChatMsgResult.TargetId) and surfaced as init-onlyChatMessage.TargetId. Init-onlyChatMessage.PeerId= the other party of aWhisper, decided on the main thread inChatService.Drainwith the game’s own rule (chat_main_vm.lua:796-807— sender, or target when the sender is self; self =IPlayerIdentity.CharId). A record with no target can’t be our own echo, so peer = sender even before self is known; with a target and self unknown it stays 0 (the plugin decides). Private HISTORY batches: theZRpcImpl.SendBytescapture also readsGetChipChatRecordsRequest.target_id(field 5, newTryParseGetChipChatRecordsRequest) keyed by call_id beside the channel, and channel-5 batches are tagged with that peer (the reply rows carry none); the batch log line gainshasPeer=(whether a peer was captured — never the charId).ChatTarget.Replyon a whisper now targetsPeerIdwhen known.ChatTarget.Whisper/WhisperTargetdocs corrected: the id is a charId, not an entity id. - System notices. The game’s “Obtained …” / item-cap / broadcast lines are CLIENT-generated (never ChitChat).
SystemNoticeService(Application) installs a Lua wrap ofZ.VMMgr.GetVM("chat_main").SetReceiveSystemMsgthat calls the original, formats withZ.ChatMsgHelper.GetSystemTypeand appends a preformatted record to a Lua queue (cap 64); C# polls every 0.25 s from theIsWorldActive-gated frameworkUpdate(one tinyDoString+ oneReadGlobalString,\x1e/\x1f-separated) and publishes through the normal chat queue:Channel=System,Type=System,IsHistory=false, no sender, framework-made NEGATIVEMsgId,Timestamp= interpolated server clock at drain. New init-onlyChatMessage.SystemLabel(e.g. “Obtained”),SystemType(E.ESystemTipInfoType: 1 item, 2 message, 3 dungeon roll),SystemRefId.Textkeeps the game’s rich-text colour tags (TMP rich text). The wrap marker lives on the VM table; each poll re-checks it and re-wraps when the VM table/method was replaced (scene reload / reconnect); a re-entrancy flag guarantees one push per notice. Logs[ChatProbe] system-notice hook installedonce (… re-installed (…gen=N)on a re-wrap). No Lua→C# callback. - Send cooldown (server-driven).
SendChitChatMsgReply {show_msg=3, int64 cd_end_time=4, err_code=5}is parsed (WireProtocol.ChitChat.Send.cs) and correlated by call_id to the channel its request named — framework sends are recorded at dispatch, the game’s chat-box sends from theSendBytesprefix (SendChitChatMsgRequest.channel_type, field 2).cd_end_timeandshow_msg.timestampare both Unix SECONDS.IChatgains two DEFAULT-implemented members (external implementors keep compiling):bool TryGetSendCooldown(ChatChannel, out double remainingSeconds)(measured against the interpolated server clock; a local Stopwatch deadline from the reply’s own stamp before the clock syncs) andevent Action<ChatSendResult> SendResult(main thread). Newreadonly record struct ChatSendResult(Channel, Ok, ErrorCode, CooldownEndUtc)+ init-onlyViaFramework/ServerTimeUtc,const ErrorTooFrequent = 5211. An accepted reply without a cooldown (Party/Guild) clears the channel’s slot; a refusal carryingcd_end_timeupdates it. Logs[ChatProbe] first send reply: …once and[ChatProbe] send refused: …per refusal. ⚠️ A plugin that CALLS these needs framework ≥2.26.0 (the defaults on an older host report “no cooldown” / never fire). - Friends. New
IPluginServices.Friends(IFriends, 7 members):Friends(FriendInfo(CharId, Name, Remark, Level, ProfessionId, GroupId, IsOnline, AvatarUrl)fromfriend_main_data:GetFriendDic(), Shield group skipped, sorted by charId),PrivateChats(PrivateChatTarget(CharId, Name, IsTop, UnreadCount, LatestText, LatestMsgType, LatestSenderId, LatestTimestampUtc, IsOnline?)fromchat_main_data:GetPrivateChatList(), game display order),event Changed, and queued actionsRequestPrivateHistory/MarkPrivateRead/CreatePrivateChat/ClosePrivateChatthat run as the game’s own coroutine VM calls (AsyncGetPrivateChatRecord(id,0,true),AsyncSetPrivateChatHasRead(id, latestMsg.msgId),AsyncCreatePrivateChat+ char-info + history likechat_service.createPrivateChat,AsyncDeletePrivateChat) insideZ.CoroUtil.create_coro_xpcall.FriendsServicepolls every 2.5 s in-world (0.75 s after an action): one tinyDoString+ two string reads; each list is rebuilt only when its serialized string changed; ids cross as decimal strings (int64-exact). Lists keep their last contents outside the world. Logs[Friends] friend/private-chat poll activeonce and[Friends] <Action> issued(no charId). - All chat/friends additions above are additive (init-only props, new types, a new aggregator property, default-implemented
IChatmembers) — binary-compatible with plugins built against ≤2.25.5. Built + unit-tested; NOT yet in-game validated. - Chat privacy in the log. The chat probe and chat service never log message text, sender names or whisper peers any more — only channel, counts and text LENGTHS (
[Chat] first received,[Chat method=3] first received,[Chat] history batch received(textLen=,hasPeer=), theSTELLAR_DIAGNOSTICSper-message lines (senderId=/textLen=),[Friends] <Action> issued). The first-correlated-Return line dumps its leading payload bytes only in diagnostics mode (they are the first history record: sender name + text). The diagnostics-only[Chat method=3 HEX]forensic dump (first 5 payloads, raw hex) is unchanged. - Local send failure. A send that never left the client (no probe,
TrySend/TrySendStickerrefused — no live chat connection, unsupported target — or threw) now raisesIChat.SendResultwith newconst ChatSendResult.ErrorNotSent = -1(Ok=false,ViaFramework=true, no cooldown, no server reply follows). Raised SYNCHRONOUSLY insideSend/SendStickerwhen called on the main thread (ChatServicelearns the main thread id inDrain), so a chat box can keep its draft instead of clearing it; queued for the nextDrainfrom any other thread or before the firstDrain. A const — inlined into callers, no runtime dependency on this build. - System notices get their own ring.
IChat.RecentMessagesis now two bounded rings merged on a global arrival number (ChatService.Ring.cs): wire chat keeps the 500-entry ring, the CLIENT notices published bySystemNoticeService(ChatMessage.SystemLabel != null) a separate 100-entry one, so a farming burst of “Obtained …” lines no longer evicts World/Guild chat or the login history. Server System lines (TopNotice/Play,SystemLabelnull) stay in the main ring. The snapshot is still ONE list in pure arrival order (notices interleaved where they arrived), rebuilt only when either ring changed. No API change. - Lazy chat background work.
SystemNoticeServiceandFriendsServiceno longer run for everyone. Notices: nothing (no wrap install, no poll) untilChatService.HasConsumers()— aMessageReceivedsubscriber or aRecentMessagesread in the last 30 s; when the last consumer goes, oneDoStringclears a new Lua flag__stlr_sys_on(the installed wrapper stays as a pass-through, its push returns at once) and drops the queue, and the poll stops; a new consumer raises the flag again with an empty queue (no stale notices published late). Friends: polls only while aChangedsubscriber exists, an action is queued, orFriends/PrivateChatswas read in the last 30 s (the first read starts it; data follows on the next in-world tick). A plugin that subscribes at load (ChatTools) starts both at once, as before. VirtualListElement.StableSlots(initbool, VariableHeight mode, default false): rowllives in pool slotl % Pool.Countinstead of slotl - first, so a one-row shift of the window (every append to a bottom-pinned list longer than the pool, every scroll step) recycles ONLY the slot whose row left — the other slots keep their content, so their texts/avatars/bubble width-solves are not redone.PlanAdmission/Repositionwalk the window by row offset and address slots throughSlotOf; consumers resolve slot→row with the new pure helperVariableListMath.StableSlotRow(first, slot, poolSize)(unit-tested). Latched heights are keyed by data index as before, so nothing is re-measured by the remap.IFriends.MarkPrivateReadnow clears the game’s badges. The queued read mirrors the nativefriend_chat_item_pc.luasetRead: runsAsyncSetPrivateChatHasRead(cid, latestMsg.msgId, CancelSource:CreateToken())ONLY whilemaxReadMsgId < latestMsg.msgId(read at run time), and on success refreshesZ.RedPointMgr.UpdateNodeCount(E.RedType.FriendChatTab, GetPrivateChatUnReadCount(), true)(the native 3rd arg was missing),CheckMainUIFriendNewMessageand dispatchesZ.ConstValue.Chat.PrivateChatRefresh— the VM itself refreshes none of them. Previously the refreshes ran even when the RPC failed and the RPC ran with nothing unread. Debounced per charId (1 s) inEnqueue: it is a server write and an RPC error < 1000 makes the proxyDisconnect.- Chat session reset at logout.
PandaChatProbe’s history/live dedup (WhisperDedup, key(senderId, msgId, send_time)) lived for the whole process — and a whisper’s key is IDENTICAL on both sides, so after character A saw a whisper live (its own echo), character B’s history copy of it in the same game session was dropped (history batch fully deduped; repro: A whispers offline B, relog as B, open the conversation). HostOnLogoutnow callsChatService.ClearSession()after theIClientState.Logoutsubscribers: empties both rings + the undrained queue + the send-cooldown slots and calls new framework-internalIChatProbe.ResetSession()(dedup + per-sender channel hints).RecentMessagesis now really “since the last login”. Plugins that cut their view at a ring entry at logout keep working (the entry is gone = everything left is newer). - Click-away popups on first mount (
WindowService):ApplySavedRectrestored the persisted visibility, and a popup’s saved flag is usuallyfalse(Entry.SetRectpersists the CURRENTVisible, and plugins place a parked popup withSetRectbeforeSetVisible(true)). So the first mount of a session re-hid the popup straight away: one frame at the old saved spot, then parked, while the plugin still thought it was open (needed an extra click).DismissOnOutsideClickwindows now never take visibility from storage (anyapplyVisibilitycaller, including a slot switch), and aSetRectissued before their mount (Token == null, previously dropped) is kept asEntry.PendingRectand applied right afterApplySavedRectin the mount, so the first painted frame is already at the requested spot. Non-popup windows are unchanged. Tests:WindowPopupMountTests. - Rich input vertical centring (
TmpRichInputField): the typed text/hint drew top-aligned underTMP_InputField(screenshot: baseline = box top + face ascender, ~4 px above centre at 13 px) although the TMP texts were set to middle-left. Text and Placeholder are nowTopLeftwith an explicitmargintop = bottom = (28 − face line box)/2, line box =(faceInfo.ascentLine − descentLine) · fontSize / pointSize · scale, recomputed on every face swap (Latin/Thai shipped face vs system CJK face). Caret follows the line. Affects onlyInputElement.RichEmojifields (legacy InputField path untouched). New Infrastructure referenceUnityEngine.TextCoreFontEngineModule(forFaceInfo). HotkeyAction.IsActive(init Func<bool>?, default null = unchanged behaviour): an “only while” gate. While false the action is dormant —HotkeyService.Tickskips its dispatch ANDSyncBlockedKeysleaves its chord out of the Rewired block set, so with “Don’t pass hotkeys through to the game” on the press still reaches the game. Gates are polled once per tick (HotkeyService.Conditional.csRefreshConditionalBlocks, only while block-all is on and a gated action exists) and the static block set is rebuilt only on a flip — never evaluated inside the hotRewired.Keyboardprefix. Dispatch evaluates the gate live; a throwing gate = inactive (rate-limited error log). Aninitproperty, not a record ctor param, so every compilednew HotkeyAction(...)keeps binding. Unit-tested (HotkeyConditionalTests).InputElement.InsertRequest(init Func<(int Seq, string Text)>?): a change ofSeqinsertsTextat the caret the field last had while focused (replacing a selection; a bound inside a<sprite=N>tag moves out of it — pureSpriteTagEdit.InsertAt, unit-tested), then focuses with the caret right after it; unknown caret (never focused / text replaced fromGet) = end. The caret is remembered per frame while focused (UGuiTextInput.Insert.cs) because the picker click deactivates the field and TMP resets its caret on deactivation; it is placed once focus is confirmed (the FocusRequest path). Written through the field setter, soOnChangereports it. TMP string-index caret behind new optionalIRichInputCaret(8-member cap onIRichInputField).RowElement.ConstrainsWindowMinWidth(initbool, default false): the row’s natural width becomes a floor on aResizablewindow’s minimum width. Per marked row,required = windowWidth + (LayoutUtility.GetPreferredWidth(row) − row.rect.width)(the delta carries all chrome/padding — nothing hardcoded); effective min =max(WindowSpec.MinWidth, ceil(max required))over shown rows, capped atMaxWidth. Measured at mount (after the forced first layout, BEFORE the saved-rect restore) and then from the window’s apply — every 5th apply (~0.5 s) or the apply after a structural change / reskin; reads cached layout only, no allocation. Enforced through the existing paths: the window’sDragResizersMin is rewritten (resize-grip clamp), andSetRectalready clamps size to it (so a saved narrower layout is widened on load and persists through the normal settle-persist); a runtime growth widens once (>0.5 unit hysteresis), sliding left if the widened edge would leave the screen. Never shrinks. Logs[Window] <id> min width from content: Non each change. ABubbleElementinside a marked row hugs its natural width uncapped (BubbleBinding.NoCap) — otherwise its parent-width cap lowers the row’s preferred width as the window squeezes and the measurement ratchets down.SetRectnow clamps size before position, so the on-screen clamp sees the final width. Non-Resizablewindows ignore the flag.- New service
IPluginServices.PlayerProfile(IPlayerProfile, Abstractions 2.26.0) with one member:void RequestAvatarUrl(long charId, Action<string?> onResult). Fire-and-forget;onResultfires exactly once on the main (Unity) thread with the fullhttps://…pngURL, ornullwhen the player has no custom photo / the request failed. It returns the URL only — the consumer owns download, disk cache andTexture2Dcreation. May be called from any thread (it only enqueues); every game interaction and the callback run on the framework tick.charIdis the same id a receivedChatMessagecarries as its sender id, so a chat consumer passes it straight through. - Sourcing (
PlayerProfileService+.Reflection.cs, Infrastructure, all by reflection — this layer has no compile-time game refs): the request isGetSocialDataRequest{ CharId = charId, Mask = 2 }(Mask bit 2 =SocialDataTypeAvatar), issued onZproxy.SocialProxy.GetSocialData(request, 0u)→Cysharp.Threading.Tasks.UniTask<GetSocialDataReply>(Social service uuid 901134105, methodId 1). The proxy is acquired through the game’s OWN accessorZservice.ServiceRegister.GetSocialProxy(IRpcCtrl ctrl, true)fedZRpcCtrl.Instance(aZSingleton, viaStellarInterop.GetSingleton) — fetch-or-create of the registered proxy, resolved fresh each request so a reconnect can’t leave a stale transport (mirrors the chat probe’s resolve-live-on-demand doctrine). The URL is read offreply.Data.AvatarInfo.Profile.Url;nullon a faulted/canceled task,ErrCode != 0, or any null link / empty URL. - The framework has no compile-time
UniTaskreference, so the returned task is POLLED on the framework Update tick via reflectedStatus/GetAwaiter()/GetResult()(UniTaskStatus 0=Pending 1=Succeeded 2=Faulted 3=Canceled) — the exact patternGameAssetsServicealready proved for the AddressablesUniTask, not anawait. The callback is marshalled withIFramework.Post. Additive (new interface + new sub-service property on theIPluginServices/PluginServices/PerPluginServicesaggregator) — binary-compatible, no plugin rebuild. - Window-element additions below are all
initproperties or a default-implemented interface member, NOT record ctor params — binary-compatible with plugins built against ≤2.25.5. See KBWindowBuilder-Patterns.mdfor each. TextElement.RichEmoji(bool): hosts the text in TextMeshPro with rich text on and the GAME’s own chat sprite asset (TmpEmojiSprites, resolved lazily once in-world), so the game chat’s 1-based<sprite=N>tags draw as emoji images. Uses the framework TMP faces (style flags still apply; one face per string, so mixed Thai + CJK shows boxes for one script). Falls back to plain uGUITextoff-game / when TMP is unavailable (worst case the literal tag shows). Window surfaces only; ignored on HUD-overlay and Shadow text.InputElement.RichEmoji(bool): builds aTMP_InputField(TmpRichInputField, behind theIRichInputFieldadapter) instead of the legacyInputField, so<sprite=N>shows as a picture while typing; one Backspace removes a whole tag (SpriteTagEdit), focusing does not select-all, submit is accepted only while editing. Falls back to the legacy field (logs[Window][emoji] RichEmoji input fell back to the plain field) if the rich field can’t be built.InputElement.FocusRequest(Func<int>?): counter token — whenever the value CHANGES vs the value captured at build, the field takes focus like a click with the caret at the END (e.g. after an emoji-picker insert). Focus is confirmed on the next tick (both field types activate in their own LateUpdate) and retried; a request on a hidden/collapsed field is dropped after ~30 frames rather than grabbing the keyboard later.InputElement.Placeholder(Func<string>?): dimmed hint text shown while the field is empty, polled with the window refresh and diffed (return a cached string). The RichEmoji field fills theTMP_InputFieldPlaceholder it already had (face picked per string, so a CJK hint doesn’t draw boxes); the legacy field gets aTextplaceholder (italic, same font/size/padding, half-alpha text colour) created only on the first non-empty hint, so fields without one build exactly as before. Ignored byTextAreaElement.InputElement.Locked(Func<bool>?): live read-only + dimmed state (text alpha 0.45) for a “busy” field, polled and diffed. StockreadOnlyon both the legacy and the RichEmoji (TMP_InputField) field, so typing/paste/delete and the default mode’s Enter-submit are blocked while focus, caret and theGet()re-seed keep working; the box keeps its size. Null = unchanged.VirtualListElement.OnAtBottomChanged(Action<bool>?) /ScrollToBottomRequest(Func<int>?):VariableHeight+StickToBottomlists only.OnAtBottomChangedreports exactly the internal stick-to-bottom pin decision (within 2 px of the bottom; content that fits = true), once initially and then on each flip (≤1 refresh late) — for a “new messages” bar.ScrollToBottomRequestis a counter token (baseline = value at build) whose change FORCES the bottom pin for a refresh so not-yet-measured rows are followed until the view settles on the true bottom.SelectableElement.OnRightClick(Action?): right-button-down over the row, from the same ticker hook asMeterRowElement.OnRightClick(honoursFrontWindowBlocks). Never firesOnClick; the left-click path is unchanged.ListElement.OnReorder(Action<int,int>?): drag-and-drop reordering of visible items (WindowInteractionTicker.Reorder.cs, ticker-polled — no uGUI drag-handler injection). ≥6 px = drag (ghost + insertion bar), else an ordinary click; firesOnReorder(from, to)with REMOVE-THEN-INSERT indices, never for a no-op; Esc or an outside drop cancels; the click after a real drag is swallowed; the enclosing ScrollRect is locked for the press. Every hit-test callsFrontWindowBlocks.- Reorder ghost fix (
WindowInteractionTicker.Reorder.cs,MatchGhostToSource/PlaceGhost): the ghost was pinned centre-on-cursor and sized fromsrc.rect.size(cell-LOCAL units, wrong when the cell sits under a scaled chain below the ghost’s Canvas parent). Now everything is in the ghost parent’s local space: size = the source’s world corners mapped throughparent.InverseTransformPoint; anchors fixed at centre, pivot = source pivot, scale 1,ignoreLayout+ any rootContentSizeFitter/AspectRatioFitterdisabled; grab =InverseTransformPoint(src.position) − ScreenPointToLocalPointInRectangle(parent, pressPoint)taken from the PRESS point (not the 6 px threshold point); each framelocalPosition = local(pointer) + grab. Placed in the same frame as the clone, before it renders. Drop target, insertion bar, clipping,FrontWindowBlocks, Esc-cancel and click-vs-drag unchanged. IWindowHost.PrewarmText(string text)— new interface member with a default no-op implementation ({ }), so existingIWindowHostimplementors keep compiling. The framework’s host queues the text (TmpGlyphPrewarm, ≤1024 texts) and rasterizes its missing glyphs ahead of layout from the window ticker’s per-rendered-frame update: ≤8 glyphs or ~3 ms per frame, face chosen by the same rule styled text uses (TmpFontAssets.FaceFor) then its fallbacks and TMP’s global fallbacks; atlas pages uploaded explicitly. Best-effort — unwarmed glyphs render exactly as before. Any thread. ⚠️ A plugin that CALLS it needs framework ≥2.26.0 (MissingMethodExceptionon older hosts) — ship them together.HitchProbe.Count(string name, int n): per-frame event counter for[Hitch]attribution, reported asname=#Non a hitch line and reset each frame (kept apart from the time buckets so a zero-ms counter is never folded intorest). Main thread only; no-op while hitch diagnostics are off. The framework’s own glyph prewarm reports its time astmp:prewarm.- TMP dynamic faces now sample at 48 pt / padding 5 instead of TMP’s default 90 / 9 (
TmpFontAssets.Sizing.cs): same pad/pt ratio (SDF spread unchanged), ~3.5× cheaper per glyph raster, visually equivalent at UI sizes; 1024² multi-atlas pages kept. The shipped faces use the sizedCreateFontAssetoverload resolved by reflection (itsGlyphRenderModelives in an unreferenced assembly) and fall back to 90 / 9 with a[Window] TMP sized …log; the system CJK / Hangul faces take the point size via the family overload (TMP picks their padding). - Variable-height
VirtualListElement: at most 8 not-yet-latched rows are active (bound + measured) at once (MaxUnlatchedRows,WindowBuilder.VariableList.Measure.cs); further unlatched slots staySetActive(false)until a latch frees budget, nearest the anchor first (bottom-up while pinned), so a login history batch or a tab switch fills in over a few applies instead of binding ~40 bubbles in one. Rows already in flight keep their admission; rows that latched before a width/count invalidation are exempt. Conditionals inside an inactive slot no longer poll or lazy-build (CondBinding.Gate). Trade-off: a long jump into never-shown history can show an empty viewport for ~0.1–0.2 s while rows latch. - Login freeze (measured 2026-10-23, worst
apply:chattools.main0.9–4.3 s → ~70 ms in-game): caused by TMP rasterizing ~60 new glyphs at pt90 inside one window apply, plus every visible row’s lazy branch building at once (and, plugin-side, synchronous avatar decodes). Fixed by the three items above together with the ChatTools plugin moving avatar decode off the apply path and callingPrewarmTextfor incoming messages. See KBChat-System-Architecture.md. RowElement.OnHover(Action<bool>?, init): pointer enter/leave over the row’s rect, from the window ticker’s existing hover list (TickHoversAndPulses—RectangleContainsScreenPoint+FrontWindowBlocks,falsewhen the row goes inactive). Registered inWindowBuilder.BuildRowonly when set; it adds no Image, raycast target, tint or padding, so a row with it lays out and draws exactly like one without. For hover-revealed controls that must not move anything (ChatTools’ per-message translate button). Rows that leave it null are untouched.TextElement.Ellipsize(bool, init): one line, cut with a trailing “…” to the text’s own laid-out width and re-fitted when the string or that width changes (TextBinding.EllipsizeToRect+UGuiPrimitives.EllipsizeFit, a binary search over the cut length so a window-resize drag stays cheap; the text also gets aRectMask2D). Styled / RichEmoji (TMP) text usesTextOverflowModes.Ellipsisinstead (StyledTextSpec.Ellipsis; a face without U+2026 hard-truncates); Shadow / HUD-overlay ignore it. For a phrase in a weighted cell next to fixed controls (ChatTools quick phrases). Texts that leave it false are untouched.TextElement.CenterGlyphs(bool, init): vertically centres the DRAWN geometry in the text’s rect instead of the font line box —StyledTextSpec.CenterGlyphs→ TMPTextAlignmentOptions.Midline/MidlineLeft/MidlineRight(Midline centres the mesh extents; the defaultCenter/Left/Rightare the Middle row, which centres ascender..descender). A lone<sprite=N>sits on the baseline, so under Middle it drew above its box centre (ChatTools’ emoji tab icon / ☺ button / emoji grid). No fixed offset → scales with any font size; preferred size (layout) is unchanged. TMP path only (styled / RichEmoji); plain, Shadow and HUD-overlay text ignore it. Texts that leave it false are untouched.SelectableElement.Bare(bool, init): a click target with no visuals — no sprite, every tint stateColor.clear(the Image stays as the alpha-0 raycast target), zero VLG padding/spacing,flexibleWidth0 so it hugs its child;SelectedBarignored. Click,OnHover,OnRightClick, FrontWindowBlocks and the reorder click-swallow are unchanged. For link-styled labels (ChatTools’ clickable whisper names). Selectables that leave it false are untouched.SelectableElement.OnClickWithRect(Action<WindowRect>?, init): fires on a left click right afterOnClickwith the row’s live rect in screen PIXELS, top-left origin (same space asButtonElement.OnClickWithRect), so a plugin can anchor a popup to the clicked element (ChatTools: channel chip, 文A language menu, “+ New chat”). Computed withRectTransform.TransformPointof the local rect corners — folds in every scale on the chain and avoids the IL2CPPGetWorldCorners(new Vector3[4])copy-out trap. Gated byClickSuppressedlikeOnClick.SetRecttakes CANVAS units: multiply byCanvasWidth / ScreenWidth. Selectables without it are untouched.BubbleElement.OutlineColorFn(Func<ColorRgba?>?, init): a per-refresh 1-px rounded ring along the bubble edge (colour = ring on, alpha honoured; null = off).WindowBuilder.AddBubbleOutlineadds anignoreLayout, non-raycast, stretchedImagechild AFTER the content child, usingWindowThemeAssets.BubbleOutline(radius)— a whiteRoundedTextureBaker.RoundedBorderedring (transparent interior, 1 px border) baked once per radius at the same radius as the fill sprite (≤0 → the 3-radius chip), kept across theme re-bakes and freed inDestroyAll.BubbleBinding.ApplyOutlinediffs the colour, so idle bubbles write nothing. Never measured → toggling it can’t resize the bubble. Bubbles that leave it null create no ring at all.VirtualListElement.RowRevision(Func<int,long>?, init;VariableHeightonly; the record moved to its ownVirtualListElement.cs— pure move): a per-data-index CONTENT revision.VariableListBindingstamps the revision a row latched at (_hRev); when a latched, active row’s revision differs,RemeasureRevisedRow(WindowBuilder.VariableList.Measure.cs) arms an in-place re-measure (deferred one apply, like theRowIdverify, because the slot’s content refresh runs afterMeasureActiveSlots), re-reads it with the same two-readStableEpsconfirm as a first latch, and replaces_honly on a confirmed change → one offset rebuild through the normalmeasuredChangedpath (bottom pin / top-row anchor unchanged). Unlike theRowIdverify it never drops the latch, so the row is NOT parked off-screen: it stays visible at its old height for the ~2 applies the confirm takes (content may overlap/underfill the slot that long), then snaps. Independent ofRowId(a droppedRowIdlatch takes precedence)._remeasureis cleared byInvalidateHeights/ResetStateand on re-latch. Null = byte-identical behaviour.- Clip-aware ticker hit-tests (
WindowInteractionTicker.Clip.cs,OutsideClip): the ticker’s manualRectangleContainsScreenPointhit-tests now also require the pointer to be inside every enabled ancestorRectMask2D/Maskof the registered cell — hover (TickHoversAndPulses: tile hover,SelectableElementhover incl.Bare,RowElement.OnHover, dropdown items), right-click (TickRightClick), raid-slot drag source/target (HitSlot/HitDropTarget), voice-click exclusion, colour-picker drag areas, render-host drag/zoom and chart pan. Before, a row scrolled half out of aScrollRectviewport still hovered / took right-clicks over its clipped part (e.g. a whisper-list row lighting up while the pointer sat on the composer below the list). The EventSystem path (Button clicks) was already correct:RectMask2DandMaskareICanvasRaycastFilters, soGraphic.Raycastrejects a press on a clipped part. The per-cell clip chain is resolved once and cached by native pointer (no per-frame allocation; cache cleared inPrune).FrontWindowBlockschecks are unchanged; the reorder path keeps its own equivalent viewport check. Cells with no clipping ancestor behave exactly as before. - (dev 2.25.5) A
ResetScroll(tab/channel switch) on the variable-heightVirtualListElementclears every row’s latched height (ResetState), so for the 2-3 applies it takesMeasureActiveSlotsto re-latch them each row falls back to the seed/running-average ESTIMATE (the chat consumer seedsRowHeight: 28f, shorter than a real wrapping bubble).VariableListMath.BuildOffsetsthen packed the rows at that too-short estimate whileRepositionsized each slot to its real content height — so adjacent rows visibly OVERLAPPED inside the clipped viewport until the real heights latched and the offsets rebuilt (“brief overlap before it shows correctly”). A burst of new messages hit the same path more briefly. - (dev 2.25.5) Fix (
WindowBuilder.VariableList.cs,VariableListBinding.Reposition): a row whose height is not yet latched (!_measured[logical]) is PARKED off-screen (newOffscreenY = 1_000_000f, a large positive y — content/slots are top-anchored pivot y=1, so a big positive y sits above the viewport top, where the ScrollRect viewport’sRectMask2Dclips it) instead of being drawn at its wrong estimate offset. The slot stays ACTIVE and is still sized to the estimate rowH so the VLG lays it out andMeasureActiveSlotscan read its preferred height (measurement requiresactiveSelf, so we hide by position, NOTSetActive(false)). As each row latches,measuredChangedbecomes true that apply →Repositionre-runs and snaps the just-latched rows on-screen at their true offsets;StickToBottomre-pins to the true bottom as offsets grow. Net: the viewport briefly blanks (~2 frames, ~33ms) on a tab switch instead of flashing overlapped rows; a measured row always takes the on-screen branch, so steady state and scrolling are byte-identical.VariableListMathand theVirtualListElementAPI are unchanged — binary-compatible, no plugin rebuild. See KBWindowBuilder-Patterns.md(VARIABLE-HEIGHT section). - (dev 2.25.4) 2.25.3 made the variable-height
VirtualListElementheight model a measure-ONCE cache to stop scroll jitter, but it latched a row’s height from the FIRST reading the slot was active for.BubbleElement’s width-solve (BubbleBinding) settles a tick AFTER that first measurement, so the firstLayoutUtility.GetPreferredHeightwas read while the bubble was still ~0 wide — the text wrapped to ~1 char per line → a far-too-tall preferred height that_measured[]then locked in permanently. The text corrected visually when the width solved, but the row’s reserved height stayed stuck tall → a large empty gap below every message. (The pre-2.25.3 re-measure-every-apply code self-corrected, which is exactly what caused the jitter 2.25.3 removed.) - (dev 2.25.4) Fix (
WindowBuilder.VariableList.cs,VariableListBinding.MeasureActiveSlots): latch a height only after it is STABLE across two consecutive measurements. New per-index_provH[]/_prov[]hold the last apply’s provisional (unconfirmed) reading; a reading commits to_h[]/_measured[]only when it matches the stored provisional withinStableEps(1.5px — above sub-pixel uGUI wobble, far below the pre-width-solve tall-vs-real gap). A pre-width-solve tall reading never matches the next (settled) reading, so it stays provisional and never locks in; a genuinely settled row matches its own next reading and latches on the 2nd measure, after which it is never re-measured — so 2.25.3’s no-jitter property is preserved (a settled row still never churns the offsets or scroll). While a row is only provisional it keeps using the running estimate in the offsets (nomeasuredChanged), so the settle window adds no offset churn._provis cleared alongside_measuredon a viewport-WIDTH change / Count DECREASE (InvalidateHeights) andResetScroll(ResetState), and grown with the cache inEnsureCapacity.VariableListMathand theVirtualListElementAPI are unchanged — binary-compatible, no plugin rebuild. See KBWindowBuilder-Patterns.md(VARIABLE-HEIGHT section). - (dev 2.25.3) The variable-height
VirtualListElement(2.25.0 measured/wrapping list — chat logs ofBubbleElementrows) re-read every visible slot’s uGUI preferred height on EVERY apply (MeasureActiveSlots→LayoutUtility.GetPreferredHeight, with aMathf.Abs(h - _h[l]) > 0.5fre-measure branch). Sub-pixel uGUI layout jitter — compounded byBubbleBinding’s width-solve running later in the same tick (ApplyValues) and marking layout dirty, so a freshly (re)bound row was measured a tick before its wrap settled — made a SETTLED on-screen row’s measured height cross the 0.5 px threshold frame-to-frame. Each crossing setmeasuredChanged→ rebuilt the cumulative offsets (VariableListMath.BuildOffsets) → re-droveContent.anchoredPosition.yviaSetScroll, fighting the ScrollRect’s own inertia. The anchor compensation only pins the topmost-visible row, so mid-viewport rows visibly shifted (“back and forth”). - (dev 2.25.3) Fix (
WindowBuilder.VariableList.cs,VariableListBinding): the per-data-index height model is now a measure-ONCE cache.MeasureActiveSlotsskips any index already in_measured[]and the re-measure branch is gone — a settled row’s height can no longer change, so it can’t churn the offsets or the scroll. A stored height is dropped + re-measured (InvalidateHeights) only on a viewport-WIDTH change (new_lastViewportWcompare — rows re-wrap → every height changes),ResetScroll(existingResetState), or a Count DECREASE (count < _lastCount— rows dropped/filtered/rolling-buffer front-dropped → indices remap); invalidation runs BEFORE the measure so active slots re-measure the same apply and only off-screen rows fall back to the estimate. An append (Count INCREASE) keeps the cache and measures only the new tail indices.measuredChanged→ the topmost-visible anchor compensation now fires only on a row’s one-time estimate→measured correction as it scrolls into view, not every tick; append,StickToBottom, and tab-switch paths are untouched.VariableListMathand theVirtualListElementAPI are unchanged — binary-compatible, no plugin rebuild. Known residual: a rolling buffer that front-drops WITHOUT changing Count remaps indices undetectably, leaving cached heights one row stale until the next width/shrink/reset invalidation (bounded, self-heals; the old per-apply self-heal is traded away to kill the jitter). See KBWindowBuilder-Patterns.md(VARIABLE-HEIGHT section). - (dev 2.25.2) 2.25.1 moved the per-request channel capture to a Harmony prefix on
ZTcpClient.Send(ReadOnlySpan<byte>)(PandaChatProbe.Send.Hooks.Tcp.cs,TryCaptureHistoryChannelparsing the framed packet). In-game that prefix NEVER fired: the one-shot[ChatProbe] Send PREFIX first invocationline (emitted on the first invocation regardless of payload) had zero occurrences, so_historyChannelByCallIdwas never populated and every login history batch resolvedbatchChannel=Unknown. Root cause: the game’s own outbound RPC does not use that 1-argZTcpClient.Sendoverload — it is a plugin/helper entry (PandaWireTap.Send.csalready documents this, and on release_3.7 only the 1-arg overload exists). Lua RPCs (ChitChat chat sends + theGetChipChatRecordshistory fetch) goZRpcImpl.ProxyCall(…byte[]…)→ProxyCallInternal→ZRpcImpl.SendBytes→ transport, and the World channel can be TCP OR UDP/KCP (ZUdpClient.Send), which a TCP-send hook could never see. NoProxyCalloverload returns the call_id (all returnUniTask<IProxyReturn>), so a ProxyCall-only hook can’t key by call_id either. - (dev 2.25.2) Fix: capture at the transport-agnostic
ZRpcImpl.SendBytes(ulong uuid, uint stubId, uint callId, uint methodId, byte[] pbData, uint pbSize, bool framePack) -> ZRpcError(newPandaChatProbe.Send.Hooks.Rpc.cs, prefixOnRpcSendBytes). It sits above the TCP/UDP split, is public + cold (PosingSendTapalready patches it in release_3.7 and proved it fires; arg layout verified UuidArg=0/MethodArg=3), has noin/refstruct param (no trampoline crash), and carries BOTH thecall_idand the raw pre-framing (uncompressed) protobuf body as plain params. The channel is parsed straight off the body argument viaWireProtocol.ParseChannelTypeFromGetChipChatRecordsRequest(the proven pre-2.25.1 source — the retiredZRpcImpl.ProxyCallhook parsed the same bytes) and storedcall_id -> channel_typein one invocation, so no thread-local/positional/cross-site correlation is needed. The call_id-keyed Return resolve (unconditional_historyChannelByCallId.TryRemove(callId)at the top ofProcessChitChatReturnPayload), the map bounding, and the_senderLastChannelper-sender fallback are unchanged. The deadZTcpClient.Sendcapture prefix was removed; that overload is still resolved asTrySend’s dispatch target and its connection-close hooks are untouched. Confirm next deploy via[ChatProbe] SendBytes PREFIX first invocation,first ChitChat outbound Call observed (SendBytes),first GetChipChatRecords channel captured, and a non-UnknownbatchChannel=per login batch. See KBChat-System-Architecture.md§7. - (dev 2.25.1) Login chat history carries no wire channel; the framework back-fills it by correlating each
GetChipChatRecordsreply to the channel that was requested. That correlation was a positional FIFO (_pendingHistoryChannels, aConcurrentQueue<int>) whose dequeue ran only AFTER the empty-batch / invalid-shape early-returns inProcessChitChatReturnPayload. So the first requested channel whose login history was EMPTY returned early WITHOUT dequeuing, leaving its channel stuck at the queue head; the next non-empty batch dequeued that stale head and was mislabeled (e.g. empty World left “World” for the following Party batch). It also broke under out-of-order RPC replies and enqueue/dequeue count drift. - (dev 2.25.1) Fix: correlate by RPC
call_idinstead of position. TheZTcpClient.Sendprefix (TryCaptureHistoryChannel,PandaChatProbe.Send.Hooks.Tcp.cs) now recordscall_id -> channel_typein a boundedConcurrentDictionary<uint,int>(_historyChannelByCallId, evict-oldest via a companion order queue, both capped atMaxPendingChatCalls) — it already had thecall_idand the framed request payload, so it parseschannel_typethere via a new zero-allocReadOnlySpan<byte>overload ofWireProtocol.ParseChannelTypeFromGetChipChatRecordsRequest(compressed payloads skipped — a tiny GetChipChatRecords request is never compressed). The Return resolves the channel with an UNCONDITIONAL_historyChannelByCallId.TryRemove(callId)at the TOP ofProcessChitChatReturnPayload, before the early-returns, so an empty batch removes its own entry (poisoning no other) and the map never leaks; a Return with no recorded id staysUnknownand falls through to the existing_senderLastChannelper-sender hint. The now-redundantZRpcImpl.ProxyCallchannel-capture hook (PandaChatProbe.Send.Hooks.cs) and the_pendingHistoryChannelsqueue were retired, removing a Harmony patch from a game-critical RPC method. The live Notify path was already correct and is unchanged. See KBChat-System-Architecture.md§7. - (dev 2.25.0)
VirtualListElementgains twoinitprops (additive, binary-compatible with ≤2.24.0):VariableHeight(measured rows) andStickToBottom(newest-at-bottom scroll). InVariableHeightmode the renderer measures each pooled slot’s uGUI preferred height (LayoutUtility.GetPreferredHeightafter a layout pass — the same fit aBubbleElement/ContentSizeFitterperforms) once it binds to a data index, keeping a per-data-index height model + cumulative-offset array; un-measured rows useRowHeightas a running-average estimate, so the scrollbar is approximate immediately and refines as rows scroll into view. The consumer drives it exactly like the uniform list (Func<int> Count, a recycledPool,Action<int> OnWindow(first), viewportHeight) — the only extra requirement is sizing the pool to cover the viewport at the shortest row. - (dev 2.25.0) Scroll stability: a measurement that replaces an estimate anchors to the current topmost visible row (its top edge is held under the viewport top) so the view does not jump; rows below reflow.
StickToBottompins to the bottom on append while at the bottom and preserves position when scrolled up. New pure helperVariableListMath(BuildOffsets/FirstIndex/MaxScroll, BCL-only, unit-tested) mirrorsVirtualListMath. Renderer:WindowBuilder.VariableList.cs(VariableListBinding, behind the newIVirtualListBindingalongside the uniformVirtualListBinding). XML-documented (CS1591). See KBWindowBuilder-Patterns.md. - (dev 2.24.0) New element (additive):
BubbleElement(HudElement Child)with init propsFillColor(ColorRgba?, alpha honoured; null = themed menu-accent default, re-tinted on theme change),MaxWidth(0 = clamp to parent/available width),CornerRadius(0 = reuse theSelectableElement/SwatchBgchip; > 0 bakes a cached rounder sliced sprite),PaddingX(10),PaddingY(6). Background is the rounded slicedSwatchBgsprite tinted toFillColor;raycastTarget = false(click-through-neutral — wrap in aSelectableElementfor clicks). - (dev 2.24.0) Content-hug:
LayoutElement.flexibleWidth = 0+ a per-applyBubbleBindingthat pins the bubble width tomin(child natural preferred width, cap). The child’s natural width is wrap-independent (uGUIText.preferredWidth), so pinning narrower never feeds back/oscillates; a verticalContentSizeFittergrows the height when the child wraps. Needs a width-respecting parent (aRowElement); dropped directly into aColumnElementit fills the column (same force-expand gotcha as a fixed-size icon). XML-documented (CS1591). See KBWindowBuilder-Patterns.md/Chat-System-Architecture.md§6.
[2.23.1] - 2026-10-21
Section titled “[2.23.1] - 2026-10-21”2.23.1 (patch) — restores the Module Optimizer after the latest game update; no API change, no plugin rebuild.
- Module Optimizer: clicking Optimize returns ranked module combos again. After the latest game update it had stopped showing any results.
Developer notes
Section titled “Developer notes”release_3.7_2added per-attribute module leveling and relocated each part’s magnitude:Mod.ModInfosis now empty at runtime; the values live inItem.ModPartLevel.Levels(MapField<int,int>, keyed by attrId — protostru_mod_part_level.proto→message ModPartLevel { map<int32,int32> levels = 1; }). No type/field was renamed (static dump-diff identical); a pure runtime data relocation, so the old positional read (Mod.ModInfos[uuid].InitLinkNums[i]paired withModNewAttr.ModParts[i]) returned 0 for every part and the optimizer scored every combo at ~0.PandaInventoryPullReadernow sources each part’s value fromItem.ModPartLevel.Levels[attrId](for each attrId inModNewAttr.ModParts), falling back to the positionalInitLinkNums[i]path for pre-3.7_2builds.ModulePartshape unchanged — only the value source moved. Unchanged:Mod.ModSlots(equipped), the pool walk, the is-a-module filter, category lookup, all public API / domain types.- Framework-only;
StellarModuleOptimizerPluginneeds no change (alreadySlotCount = 5, consumes onlyIInventory). Validated in-game on liverelease_3.7_2. PR #129.
[2.23.0] - 2026-10-19
Section titled “[2.23.0] - 2026-10-19”2.23.0 (minor) — mods can mix sound into your party voice chat with your mic; chat no longer crashes or drops a message after a reconnect; and the layout editor gains a draggable toolbar with per-plugin filtering, arrow-key nudging and show/hide. Adds API for plugins (IVoiceInjection, Abstractions 2.23.0); additive, no plugin rebuild.
- Mods can now mix their own sound — a file or generated speech — into your party voice chat alongside your microphone, so teammates hear both. Several can play at once.
- The layout editor has a new toolbar you can drag anywhere, with drop-down menus, filtering to one plugin’s windows, and arrow-key nudging to place things exactly.
- You can now show or hide individual overlay elements right from the layout editor.
Changed
Section titled “Changed”- “Reset” in the layout editor now also brings back anything you had hidden, and is split into “Reset filtered” and “Reset all”.
- Windows like Settings, the data inspector and the combat-meter history now keep one position across all your layout slots instead of moving per slot.
- Sending a chat message right after the game reconnects (switching lines, changing scenes, or recovering from a dropped connection) no longer crashes the game.
- Chat messages sent immediately after a reconnect now go through instead of silently failing until the next message you received.
- Fixed a multi-second freeze the first time settings were saved or you entered layout edit mode.
Developer notes
Section titled “Developer notes”- New API (additive):
IPluginServices.Voice→IVoiceInjection(Available,Status,LoadFileAsync(path)WAV natively / MP3/WMA/M4A via Media Foundation,LoadPcm(ReadOnlySpan<short>, rate, channels)for TTS/generated audio,CreateSource()→IVoiceSource?,SetUplinkPolicy(disableNoiseSuppression?, disableAgc?, disableVad?, bitRateBps?));IVoiceSource(Clip/Volume/Loop/Playing/Position,Play/Stop/PlayOnce);IVoiceClip(Duration); theVoiceInjectionStatusrecord. XML-documented (CS1591). - Backed by a process-wide native DirectSound capture hook on
GCloudVoice.dll(IAT slot+0x6190D8= DSOUND ordinal 6DirectSoundCaptureCreate→CreateCaptureBuffer+0x18 → capture-bufferLock+0x40), promoted intoStellar.Infrastructure.Gamefrom the in-game-validated StellarExperimentPlugin single-clip prototype and generalised to a fixed 16-source allocation-free mixer. TheLockhook runs on GVoice’s native capture thread (no alloc/lock/log/IL2CPP, whole-body try/catch, always returns the original HRESULT). Installed at boot fromBootstrapPlugin.ApplyLifecyclePatches(title screen, before login’sGCloudVoice_Init), mirroringInstallInstrumentToneRelay; aProbeInstallfallback (opens the mic briefly) covers hot-reload late attach and is surfaced viaStatus, not the public API. Clips keep their source PCM and are re-resampled to the live capture format off-thread; the uplink DSP (GCloudVoice_invokeNS/AGC/VAD +GCloudVoice_SetBitRate) is applied while any source plays and restored when all stop, driven from a main-thread maintainer on the framework Update tick. Sources a plugin creates are auto-disposed on plugin teardown (PluginVoiceInjection, mirroringIHarmonyHost). DLL offsets are pinned toGCloudVoice.dll 4.3.0.253999397; guards fail safe (no patch, logged) on a version change. Seedocs/design/framework-2.23-voice-injection.mdand KBVoice-Chat-GVoice.md. - Chat:
IChat.Sendresolves the live chatZTcpClienton demand at send time viaZSingleton<ZRpcCtrl>(Gateway/World RPC transport); the passive packet cache +Close/CloseAndDelayDisposehooks are kept as a fail-soft fallback. A stale freed client can no longer be used (was a DEBUG-onlyCheckValid→ access violation) and the post-reconnect send window is closed. No API change, no plugin rebuild. PRs #126, #128; KBChat-Send-Connection-Lifetime.md. - Layout editor: toolbar migrated to a draggable uGUI window on a dedicated root canvas (capped at 32766); the per-plugin filter gates outline/select/drag/resize; dropdown controls (slot, selection readout, filter); arrow-key nudge with the game keyboard gated via
KeyboardInputGate; per-element show/hide; Reset split into filtered/all and now restores visibility; non-HUD free-drag windows (Settings/DataInspector/launcher/CombatMeter history) persist in a global slot-independent bucket; config disk writes moved off the game thread (cfg-save3378ms → <2ms) and the toolbar is pre-warmed to remove the first-edit-enter freeze; strings localized en/fil/id/ja/ko/th; binary-compatible (init props / internal interfaces / optional params). PR #127.
[2.22.1] - 2026-10-12
Section titled “[2.22.1] - 2026-10-12”2.22.1 (patch) — clicking a mod window’s buttons no longer steals your keyboard; no API change, no plugin rebuild.
- After clicking a button or row in a mod window, WASD moves your character again instead of typing into the chat box.
Developer notes
Section titled “Developer notes”- Every window click path now calls
ClearSelectionAfterClick()(already used by the base Button/Toggle):BuildSelectable,BuildTile+ pin-star toggle, the dropdown blocker, both chrome close buttons, the cooldown-tile clicker, and the meter voice/debuff cells. A clicked uGUIButtonused to stayEventSystem.current.currentSelectedGameObject, and the game routes keyboard input to UI/chat while a uGUI object is selected (the launcher rail already handled this viaClearRailSelection). - Window Buttons also get
Navigation.Mode.None(DisableKeyboardNav) against arrow/d-pad focus; that alone does not stop pointer-click selection.UGuiTextInputis untouched — text fields legitimately hold focus. PR #124.
[2.22.0] - 2026-10-10
Section titled “[2.22.0] - 2026-10-10”2.22.0 (minor) — fixes window text that could occasionally turn into garbled letters, keeps centred / edge-anchored windows in place across resolutions and width changes, and lets a free camera leave clicks on the game’s own windows to the game (IInputShield.IsPointerOverGameUi, Abstractions 2.22.0; additive, no plugin rebuild).
- Plugins with a free camera can now let you use the game’s own windows and sliders while the camera is on.
- Fixed window text (close button, HUD-style overlays) occasionally showing garbled letters.
- Windows that sit at the top centre, the middle or a screen edge now stay there at every resolution and after a plugin update changes their size (they could drift off centre).
Developer notes
Section titled “Developer notes”- Anchored window layouts (
WindowAnchor≠TopLeft): a save also stores the offset of the window’s anchor point from the canvas anchor point in design units (slots.<n>.windows.<id>.<res>.anc/.ax/.ay, beside the unchangedx/y/w/h— rollback-safe), and every restore re-places the window from that offset at the current canvas, UI scale and size (AnchoredPlacement,WindowService.Anchor.cs). A legacy exact-resolution save keeps the anchor point it had. Root cause: a Top-anchored window kept its saved absolute left edge when its width changed (Photo Studio’s free-camera HUD 800 → 1240 px sat 220 px right of centre at 2560×1440). Pinned inWindowAnchorLayoutTests(5 resolutions × 3 UI scales). IInputShield.IsPointerOverGameUi: oneEventSystem.RaycastAllper rendered frame (only when asked); true when the TOP-MOST hit sorts below Stellar’s band (sortingOrder < 32749), so a Stellar window over a game window answers false. WithSTELLAR_DIAGNOSTICS=1every flip logs[FreeCam] pointer over game UI=yes|no hits=N top=<object> order=<n>. Hidden game HUD may still answer the raycast — consult it only while the game interface is shown.WindowRenderer.OnFontTextureRebuiltno longer filters toMenuFont: it refreshes, on every live window, the legacyTextbound to whichever font rebuilt (WindowToken.RefreshFontTexture(Font)). HudOverlay text intentionally stays onConfigureText’s builtin font (native-HUD look), and that font’s atlas repacks were previously ignored → stale glyph UVs.- Window chrome close buttons (✕, title-bar and overlay-corner) and the legacy-
Texttitle fallbacks (title bar, Tracker/Party overlay) now getMenuFontat build time instead of only on the first theme switch. - New rate-limited (≤1/s)
[Window] font atlas rebuilt: …log line: font name, atlas size, Text refreshed, suppressed count.
[2.21.0] - 2026-10-09
Section titled “[2.21.0] - 2026-10-09”2.21.0 (minor) — Korean UI language. Adds "ko" as a language code plugins can receive and WindowSpec.TitleProvider for live-localized window titles (Abstractions 2.21.0); additive, no plugin rebuild. Plugins can also see where another player is (zone, line and position) once their profile card has been opened (SocialSnapshot.Location, additive).
- Korean (한국어) is now a language option — pick it under Settings → Themes → Language, or keep “Follow game client” on a Korean game client.
- The Stellar launcher’s Plugins and Settings labels now follow your language.
- More of the Stellar settings and theme editor now follow your language.
- Pinned launcher tiles stay pinned after plugins update.
- Hotkey names in Settings now follow your language.
- Plugin names in Settings now follow your language.
- Plugins can now show where another player is — their zone, line and spot on the map — once you’ve opened that player’s profile card, even when they’re far away or in a different zone.
- Follow game client now picks your game’s language reliably (it could fall back to English).
- Coloured text on HUD bars, windows and notifications is no longer blurred by its shadow.
- Hotkeys with no key set now say so in your language.
- Turned-off plugins keep their proper name in Settings instead of a file name.
- Twin Striker gear details (advanced stats) now show correctly.
Developer notes
Section titled “Developer notes”- New
kocatalog code; plugins shipLang/ko.json(embedded as<Ns>.Lang.ko.json). The framework’s own UI ships a full Korean catalog (Stellar.Infrastructure/Lang/ko.json). - Supported languages now come from one list (
UiLanguages) shared by the engine, resource scan and Settings dropdown; a test fails the build if a listed language lacks a complete catalog. - The client-language mapping (
ClientLanguageLatch.SupportedLanguage, formerlyClientLanguageProbe) mapsLanguageType.ko = 4. Hangul font coverage: Noto Sans CJK KR + Malgun Gothic appended to the END of the OS-font chain (OverlayFontFamilies.Chain, used byWindowThemeAssets), and Malgun Gothic is attached as a TMP fallback asset of the chosen CJK face (TmpFontAssets) — Yu Gothic UI / Meiryo carry no Hangul, so on real Windows Korean styled text would otherwise render as boxes. On Proton, Source Han Sans already covers Hangul. - Pre-release font regression caught in owner testing (never shipped): the first Korean pass inserted Noto Sans CJK KR + Malgun Gothic BEFORE the Latin tail of the OS-font chain.
Font.CreateDynamicFontFromOSFonttakes each glyph from the first family that has it, and Malgun Gothic also carries Latin — so on a Proton prefix with Liberation Sans + Arial + malgun.ttf (no Noto/DejaVu) every English glyph switched from Liberation Sans to Malgun Gothic, and on real Windows Malgun would have replaced Arial. The Hangul families now sit after “Arial”, so Latin/Thai/Japanese resolve exactly as in 2.20.0 and only glyphs no earlier family has (Hangul) fall through. The chain moved into a Unity-freeOverlayFontFamiliestype soFontFamilyOrderTestscan pin both rules (every Latin family precedes every Hangul family; the 2.20.0 chain is an unchanged prefix). The TMP path (TmpFontAssets) is unaffected: Latin styled text uses the shipped Noto/Thai faces, and Malgun Gothic is only the last CJK candidate / a Hangul fallback asset of the CJK face. LauncherView’s “PLUGINS” section label and “Settings” tile were hard-coded English in every language — missed by the Korean pass because they don’t go through a Settings panel. Fixed:LauncherViewnow takes anILocalization text(same pattern asThemesPanel/PluginsPanel), resolved each frame via_text.T("launcher.plugins")/_text.T("launcher.settings"); wired fromWiring.Launcher.csusing the existing_frameworkLocalization. New keys in all six catalogs.- Remaining framework hard-coded English, found from an in-world Korean check: the Settings hub window title + its Shift+Home hotkey description (
settings.window.title/settings.hotkey.toggle, resolved inWiring.Settings.csvia_frameworkLocalization); the theme-editor’s colour-slot labels (Accent, Panel background/accent/border, Warning, HUD accent —theme.color.*), Edit/Reset/Remove buttons and the two name-validation errors (ThemeEditorBody/.Describe.cs); the Themes tab’s button/scrollbar style names (Outline/Filled/Glass, Thumb/Track —themes.style.*), preset names (Default/Dark/Light/Crimson —themes.preset.*, display label only, never the persistedThemePresetvalue) and the Stamina preview caption (ThemesPanel.cs; “HP” stays a kept symbol); the Shift+` layout-edit toolbar (heading, filter chips, Slot/Selected/Reset selected/Exit —layout.toolbar.*, reusinghotkeys.filter.all/tab.gameui/common.resetAllwhere the text already matched) and the Settings → Game UI list + its edit-mode outline label (10 native-UI element names —gameui.entry.*).FrameworkColorRegistration.RegisterAllruns inLoad()before the hot-update assemblies (and so the real “follow” client language) are resolvable, so it still registers the English fallback label; a newColorRegistryService.Relabel+FrameworkColorRegistration.RelabelAllre-applies the correct label, andThemeEditorBody(not Host, not a per-tick poll) owns WHEN: the first relabel is folded into its own slot-list cache refresh (SlotAt), which runs only when this editor’s element-tree Funcs are actually pulled byWindowService— a pull that is skipped entirely while the Settings window is hidden (if (!e.Visible) { ParkOrDestroy(e); return; }, no value-pull at all) — so this relabel’s first client-language read happens only after the player has opened Settings, never merely “N ticks after boot”; the second trigger is a later explicitLanguageChangedswitch. (An earlier pass of this fix polled it from every framework tick regardless of window visibility, which still read the client language before the game finished its own init — caught in owner review and corrected.) A newIThemeOverrides.Revisioncounter (bumped by Register/Unregister/Relabel) fixes the editor’s slot-list cache, which was keyed onSlotCountalone and so missed a Relabel (the count never changes); a small pureSlotCacheRefresh.ShouldRefresh(count, revision, cachedCount, cachedRevision)helper inStellar.Applicationpins that decision with a test the inlineThemeEditorBodycheck alone couldn’t get (Unity types block unit-testing that class directly). The row prefix for the framework’s own colour slots (“Theme · Accent”) is display-mapped to the active language (theme.owner.framework) without touching theOwneridentity field the picker/system-slot checks compare against.LayoutEditorOverlaytakes itsILocalizationvia a newSetLocalizationsetter (its constructor is already at the 6-dependency cap), wired fromWiring.Settings.csalongsideSetNativeUi/SetWindows.NativeUiAllowlistEntry/NativeUiEntryDescriptornow carry aDisplayNameKey(agameui.entry.*catalog key) instead of literal display text;GameUiPaneland the layout-editor’s native-UI outline label resolve it throughILocalizationat render time. (The earlier known limit — Settings hub title and hotkey descriptions baked once at wiring — is fixed: see theWindowSpec.TitleProviderand framework-hotkey relabel notes below.) Left English (reasons in the task report): the perf-harness overlay (stellar.perf-overlay, dev-only diagnostic, gated onPerfProbe.IsEnabled); the Themes-tab live preview’s decorative sample data (“Lv 78”, “Ribery / Wind Knight” — mock character name/class, same category as its neighbouring numeric placeholders); the “Stellar” wordmark and the ✕/▾ glyphs (already-established intentional exceptions). - Framework hotkey descriptions follow the language: the five actions the framework declares for itself (
framework.layout-edit,framework.hud-toggle,framework.hud-hold,framework.perf-toggle,framework.settings-toggle— ids unchanged, now constants inStellar.Application.Services.FrameworkHotkeyIdsbecause they are persisted config keys) are declared during wiring, before hot-update /Game.Init, so they keep an ENGLISH description at declaration —DeclareSettingsHotkeypreviously calledT("settings.hotkey.toggle")there, an early client-language read inOnHotUpdateReadythat could latch the wrong language index for the session; it now declares English too. (The hub’s window TITLE was a second eager read in the sameWirePhase9Ui— see theWindowSpec.TitleProvidernote; with both fixed, no framework wiring path reads the client language before Settings is shown.) A new internalHotkeyService.Relabel(id, description)replaces only the description (binding, suggested default, callback and persisted config untouched; unknown id = no-op;IHotkeyAction.Descriptionis read live by the panel, so no cache/revision is needed), andStellar.Infrastructure.UI.SettingsPanels.FrameworkHotkeyLabelsmaps each id to a catalog key (hotkey.framework.*+ the existingsettings.hotkey.toggle).HotkeysPanelowns WHEN, same pattern as the theme-colour labels: the first relabel is folded into its action-row label Func (RowLabel), whichWindowServiceonly pulls while the Settings window is shown with the Hotkeys tab branch active (hidden windows are parked with no value-pull;Conditionalbranches build lazily), and again on everyILocalization.LanguageChanged. Nothing on a per-tick path (PollCaptureUgui,HotkeyService.Tick) reads the language. The perf-overlay toggle’s description is included because it costs one map entry. Pinned byHotkeyRelabelTests(Relabel touches only the description; unknown id no-op; the map covers exactlyFrameworkHotkeyIds; every key exists in all six catalogs; ids equal the shipped persisted values). WindowSpec.TitleProvider(Abstractions, additive) +WindowSpec.DisplayTitle: an optional live-localized display title, mirroringLauncherEntry.TitleProvider(Titlestays the fallback; identity isId). The window chrome (WindowBuilder.BuildTitleText, and the tracker/party overlay chromes) reads it when it BUILDS the window — whichWindowServicedoes only once the window is first shown — and the title-bar and tracker/party overlay chromes re-read it in a reskin action, which runs onInvalidateTheme(theme change andLanguageChanged), never per frame; a throwing provider falls back toTitleand is logged once per window id.docs/plugin-development.md§ Localizing now tells plugin authors to localize window titles / launcher tiles viaTitleProviderand never callT()at registration. The Settings hub now registers with Title “Stellar Settings” +TitleProvider→settings.window.title(SettingsHubSpec, Infrastructure), removing the last eager client-language read inOnHotUpdateReady(RegisterSettingsHub’sT("settings.window.title")) and letting the hub title follow a live language switch. The hotkey relabel trigger moved intoFrameworkHotkeyRelabeler(constructed byHotkeysPanel, which is not unit-constructible — its capture partial loads UnityEngine). Pinned byRenderDrivenRelabelTests(counting fakeILocalization: building the hub spec and the relabeler performs noT(); first row render relabels once;LanguageChangedrelabels again).- Plugin names in Settings (Plugins tab, Hotkeys group headers, Performance → per-plugin update rate) are display-mapped to the plugin’s own localized launcher tile: new internal
PluginDisplayNames(IPluginDisplayNames, Application) returns theDisplayTitleof the FIRST launcher entry that plugin registered, else its internal name. Ownership is recorded at registration: a newPerPluginLauncher(same decoration asPerPluginHotkeys, wired throughPerPluginScope.LauncherbyPluginHost) forwardsILauncher.RegistertoLauncherRegistryvia the internalILauncherOwnedRegistrationssink tagged with the plugin guid (per-INSTANCE, reference-keyed:LauncherEntryis a record).LauncherRegistry‘s dispose now removes by reference, so disposing one of two value-equal entries can no longer evict the other plugin’s. Resolution is render-driven only (the panels’ row/header Funcs, whichWindowServicenever pulls for a hidden window); a throwingTitleProviderfalls back to the internal name, logs once, and is not invoked again for that plugin. A plugin that disposes its launcher tile when disabled keeps the last title it showed while disabled, until a language switch (then its remembered declared name — see the disabled-plugin note below). DISPLAY ONLY: ids, enable/disable, config keys, rate prefs, logs, collapse state and sort order all stay on the internal identity (Hotkeys groups still sort by plugin guid, the Plugins/Performance lists keep registration order) — so a list does not reshuffle when the language changes. The panels take the resolver as an init property (DisplayNames), not a constructor argument, keeping their constructors under the STELLAR0004 cap;WirePhase9Uibuilds them in a newBuildSettingsPanelshelper (STELLAR0002). Not covered: theme-editor colour-slot owner prefixes (the owner there is a plugin-chosen key prefix, not plugin identity). Pinned byPluginDisplayNamesTests. - Eight released plugins (AutoFishing, PhotoStudio, RaidManager, Wardrobe, CustomProfileImage, AutoGather, MinimalNameplate, LoadoutSwitcher) used to pass a TRANSLATED string as
LauncherEntry.Title; their new builds (this i18n train) pass a fixed EnglishTitle+ aTitleProviderfor display instead.LauncherRegistry/LauncherPrefspersist pins keyed byTitle(LauncherRegistry.cs,LauncherPrefs.cs), so a ja/th/id/fil user who pinned one of these under its old translated title would silently lose the pin on update. Fixed:LauncherRegistry.Registernow calls a newLauncherPrefs.MigratePinIfNeeded(entry.Title, entry.DisplayTitle)— a pureLauncherPinMigration.Migratedecides whether the saved pin set holds the OLD displayed title but not the newTitleand, if so, rewrites that one entry in place (same position) to the newTitle; a no-op (no allocation) in the steady state, via twoHashSet.Containschecks before anything is converted to an ordered list. If a user had somehow already pinned under both the old and new title, the stale old one is dropped rather than leaving a duplicate. Readingentry.DisplayTitle(which invokes the plugin’sTitleProvider) synchronously insideRegisteris deliberate, not a new early read: it reproduces the exact moment in plugin load the OLD build read its (translated)Titleat, which is what makes the saved key match, and everyTitleProviderin the current catalog is itself just_loc.T(key)already resolved no later than this point by the plugin’s ownRegistercall. Because it runs plugin code,LauncherRegistry.Registernow takes anIPluginLog(fail-safe likeWindowService.SafeApply’sShouldRenderguard): a throwingTitleProvideris caught, logged once, and only skips that entry’s migration — it can never abort the plugin’sRegistercall. A collision guard skips migration entirely when the stale displayed title is actually another already-registered entry’s ownTitle(not reachable with today’s catalog — every entry has a uniqueTitleandDisplayTitlein every shipped language — but a future name clash must not silently steal or drop a DIFFERENT plugin’s pin). Pinned withLauncherPinMigrationTests(pure helper) andLauncherRegistryTests(end-to-end throughRegister, including a reload-from-config check that the migration was actually persisted, a throwing-provider case, the collision guard, and a steady-state “noSave” assertion). - Follow-game-client latch fix (pre-existing since at least 2.20.0):
PandaClientLanguagecached the FIRSTLocalizationMgr.CurrentLanguageTypeIndexread ≥ 0, but that static’s backing field reads0(zh_Hans) until the game’sLocalizationMgrinitialises — after the framework wires its UI — so a wiring-timeT()read latchedindex=0 designLanguage=Truefor the session (measured on an en_TH client and a ja client):followserved English and the ChineseNameDesignfallback was armed. NowStellar.Application.Services.ClientLanguageLatchowns the decision: reads are a field read that never touches the game and never latch (unknown →"en",IsDesignLanguagefalse); a non-zero value latches whenever seen (only the game writes one); a0latches only on a game signal — a postfix on the game’s own setter (set_CurrentLanguageTypeIndex+SetLanguage, which also tracks an in-game switch) or theGame.OnLoginbackstop (character select). A signal only records the index and sets a pending flag (it runs inside the game’s setter postfix); the framework tick’s un-gated drain (ClientLanguageLatch.DrainPendingChange, a flag check) raisesIClientLanguageProbe.Changedon the main thread, andLocalizationEngineraisesLanguageChangedwhenfollow’s resolved code actually changes — never from aT()read or inside game code. BothLanguageChangedraise paths (Settings switch, client change) share one per-handler-guarded helper.PandaGameDataProbereads the latch through the read-onlyIClientLanguageInfoport.PandaClientLanguageis now only the rawIClientLanguageReader. Verified on the test client: hook-install readraw=0, thensource=game-setter raw=11right afterGame.Init, before the title screen UI. Pinned byClientLanguageLatchTests. SocialDataReadernow parsesSocialData.scene_data(field 10) instead of skipping it:map_id,scene_layer,line_id,pos,scene_area_id,level_map_id,level_pos,level_area_id(field 19, signed). New float readerWireProtocol.TryReadFloat(wire type 5) inStellar.Wire.- New
SocialSnapshot.Location(SocialLocation?) — an init-only property, not a positional ctor parameter, so plugins compiled against the oldSocialSnapshotconstructor still bind.null= the reply carried noscene_data; a present-but-zeroed value is kept distinct.SocialLocationaddsReceivedAtMs(server-now ms when the reply arrived; 0 pre-sync) andCurrentSceneId/CurrentPos/CurrentAreaId, which prefer thelevel_*trio (observed in-game:level_map_id/level_posare live,map_id/posstay at the last saved world spot, e.g. while the player is in a dungeon).Position.dirwas always 0 in-game, so there is no facing for far players. - Only full-mask replies (the ID card fetch, mask 0) carry
scene_data; nameplate/avatar thin-mask replies do not.SocialDataCachetherefore carries the previousLocationforward when a newer reply for the same char has none (a present-but-zeroed one still replaces it), and stampsReceivedAtMsfrom the framework’s interpolated server clock (injected inWiring.Core, read lazily).IEntityDetail.RefreshSocialSnapshotalready requests mask 0, so it returns a location too. - Diagnostics:
[SocialLocation] char=… name=… present=… map=… layer=… line=… pos=(…) dir=… area=… levelMap=… levelPos=(…) levelArea=…, logged byPandaSocialDataProbeonce per change per char. - Tests:
SocialDataReaderTests(present / empty / absentscene_data, field 19,Current*) andSocialDataCacheTests(carry-forward, stamping, zeroed replaces, no clock → 0). - Settings → Hotkeys: the unbound cell (
[ unbound ]) resolveshotkeys.unbound(new key, all six catalogs);HotkeysPaneldrops its per-action binding-label cache onLanguageChanged. Disabled plugins no longer list under their assembly name:PluginRegistryremembers each plugin’s declaredNamefrom its last successful enable (pluginssection, keyname.<id>, written only when it changes; a save failure only logs) and seeds a later boot’s disabled slot with it;PluginDisplayNameskeeps the last localized tile title it resolved for a plugin whose tile was disposed by disabling it (ForgetRememberedonLanguageChanged— the remembered title is in the old language); a plugin never constructed on this install showsPluginNameFallback.FromAssemblyName(StellarMahiruUtilityPlugin→Mahiru Utility). Display only — ids, config keys and enable state are unchanged. Pinned byPluginRegistryDisplayNameTests/PluginDisplayNamesTests. ProfessionSpecs.TalentSchool: Twin Striker’s two specs (30001 Formless, 30002 Crimson) mapped to talent-school ids 124/125 —data/StarResonanceData/tables/TalentSchoolTable.json(release_3.7) has no row 124, and 125 is “Hand Cannon” (an unrelated school). Fixed to the real ids, 128 (“Formless Spec”) and 129 (“Crimson Spec”); every other sub-profession → school mapping was checked against the table and is correct. This table feedsEquipAttrSchoolLib.TalentSchoolId-gated gear rolls (GameDataEquipService/PandaGameDataProbeadvanced-stat resolution), so a Twin Striker’s spec-dependent advanced gear rolls were being matched against the wrong (or a nonexistent) school. Pinned byProfessionSpecsTalentSchoolTests(all 18 known sub-profession ids against an embedded release_3.7 table snapshot, plus a local-only cross-check against the live JSON when the data submodule is present).
[2.20.0] - 2026-10-08
Section titled “[2.20.0] - 2026-10-08”2.20.0 (minor) — plugins can hide everything the game’s own photo screen can hide, and use the [, ] and backslash keys as hotkeys. Adds API for plugins (Abstractions 2.20.0); additive, no plugin rebuild.
- Photo plugins can now hide each thing the game’s own photo screen lists on its own: you, your Spirit Echo, other adventurers, friends, party, guild, non-players, enemies, weapons, collectibles and other players’ Spirit Echoes.
- Hiding friends, your party or your guild now really hides them, even while other adventurers stay visible. A hidden group always wins: a guild mate who is also your friend is hidden when you hide your guild. Party members stay visible while you show your party.
- Hiding weapons now hides every player’s weapon, not only yours, and they stay hidden when a weapon is swapped or a player comes back into view.
- The [, ] and backslash keys can now be used as hotkeys.
Developer notes
Section titled “Developer notes”- New
VisibilityLayers(one bit each, throughISceneVisibility.Hidelike every layer):SelfCharacter(1024),OwnSpiritEcho(2048),Strangers(4096),Friends(8192),Party(16384),Guild(32768),NonPlayers(65536),Enemies(131072),Weapons(262144),Collectibles(524288),OtherSpiritEchoes(1048576); setsVisibilityLayerSets.PlayerGroups(the four player groups) andVisibilityLayerSets.World(every world layer). Each drives the game photo screen’s ownCameraFrameCtrl.SetEntityShowtype, transcribed from its show list (camerasys_data.luaShowEntityCfg): Oneself 1, SelfPet 14, Stranger 6, Chum 2, Team 3, Union 4, FriendlyNPCS 7, Enemy 9, WeaponsAppearance 10 (the game’s Weapon toggle never writes AngelWeapons 8), Collection 12, OtherPet 15. - Legacy layers are unchanged in value AND meaning:
OtherPlayers= the master switch 11,OtherPlayers | KeepParty= Stranger + Chum + Union,Self= Oneself + SelfPet. Plugins built against 2.16–2.19 behave exactly as before. A new plugin can split “Me” intoSelfCharacter/OwnSpiritEcho, and replaceKeepPartywithStrangers | Friends | Guild. - The game shows a player while ANY group they belong to is shown, and
OtherPlayers(11) is checked before the groups: while it is hidden no group brings anyone back. Anyone no shown group rescues falls back toStrangers(Nearby), so hiding all fourPlayerGroupshides every other player (static RE ofZEntityMgr.isCharVisibleBySource). - Per-character hides (the game’s switches cannot express them):
Friends/Party/Guildhide a player who is in ANY hidden group (hide wins) — even whileStrangersis shown, and even when another group they belong to is shown; the one exception is a party member, who stays visible whilePartyis shown;Weaponshides every other player’s weapon on top of the game’s own (local-only) switch. Body =ZEntityHelper.SetModelRenderVisible(entity, false, (EVisibleSource)28, withSummon:false, withEffect:true)— a private bit of the per-source mask no game path writes; weapon =ZModelHelper.SetVisilbe(model, 0x6000, false, EModelAlphaSourceType.EPhoto)(every game EPhoto writer targets the local player only). Event-driven: one pass overZEntityMgr.CharIdListon a request change / re-assert / membership change (party roster,LuaDataMgrfriend + guild caches), a postfix onZEntityMgr.checkIsRenderVisiblefor players entering view and onWeaponModelComp.onWeaponModelLoadedfor rebuilt weapons (re-applied next tick); hooks install on first use. Release shows back only the characters the framework hid (CharacterHideLedger). Rules:CharacterHideRules(truth table pinned inCharacterHideRulesTests). EntityShowPlan.TryHideTypeFornow maps the five counters measured on the owner’s client and read from the native switch table: FriendlyNPCS 7 → Npc 1, Enemy 9 → Monster 2, Team 3 → Team 4, Collection 12 → Collection 9, OtherPet 15 → OtherPet 13 (a counter reset under a held hide is now re-hidden for them too). Me (Oneself 1) and Weapon (WeaponsAppearance 10) are per-entity on/off flags, not counters (EntityShowPlan.IsFlagType): every re-assert (game photo screen closed, scene entry, local player re-created or its weapon rebuilt) re-writes them (RewriteHeldFlags), idempotently. New diagnostics lines:[PhotoVis] flag re-assert types=[…],[PhotoVis] characters pass=…,[PhotoVis] character entered …,[PhotoVis] weapon rebuilt …,[PhotoVis] local player rebuilt …,[PhotoVis] membership changed ….- Internals: the visibility backend now drives every world layer through ONE
EntityShowPlanholding the union of the requested layers’ camera types (EntityShowTargets), so two layers sharing a type (Self+SelfCharacter) hold it once; a layer is reported inHiddenonly when every type it needs is held. The party refresh (ForceRefreshCharVisible) now runs while any player-group type is held without the master switch. The diagnostics line[PhotoVis] holds after Self …became[PhotoVis] holds after entity write target=[…]: …and covers every world layer. - New
IHotkeys.IsActionHeld(string actionId)— the level query the framework’s own hold-to-hide-HUD already used, now for plugins (poll it fromUpdatefor repeat-while-held). Default-implemented (returns false) so existing test doubles still compile. - New
StellarKeyCode.LeftBracket(91),Backslash(92),RightBracket(93), UnityKeyCodevalues; the input gateway polls them like every key, and a chord captured in Settings → Hotkeys now saves and displays by name. - Dev tooling:
tools/install-stellar.shnow copies DLLs withcp --remove-destination(a new inode), so a deploy into a running client never rewrites the copy the game has mapped.
[2.19.1] - 2026-10-05
Section titled “[2.19.1] - 2026-10-05”2.19.1 (patch) — expired potion and food buffs leave the meter again. No API change, no plugin rebuild.
- Potion and food buffs now disappear from the CombatMeter’s Buffs & Debuffs icons when they run out. In 2.19.0 they could stay there after expiring.
Developer notes
Section titled “Developer notes”BuffEffectSyncReader: aRemoveLayer(6)that leaves no layers (no payload, or a payload whoseLayeris 0 / absent) now removes the buff; one that leaves layers is still an upsert (the 2.19.0 re-eat time extension). 2.19.0 upserted EVERY payload-carrying RemoveLayer, but the consumable’s final expiry also arrives as RemoveLayer + payload with no layers left, so the expired buff stayed inICombatLookup.BuffsForforever. Pinned byBuffEffectSyncReaderTests(re-eat keeps, zero-layers removes).- Diagnostics: with
STELLAR_DIAGNOSTICS=1, every rawBuffEffectis logged as[CooldownBar][diag] raw entity=… type=… uuid=… payload=… payloadUuid=… layer=… dur=… create=…(BuffEventBatch.Trace, collected only when diagnostics are on).
[2.19.0] - 2026-10-05
Section titled “[2.19.0] - 2026-10-05”2.19.0 (minor) — one-line and multi-line text boxes for plugin windows, plus per-row bar colours, and re-eaten potion/food buffs no longer vanish. Adds API for plugins (Abstractions 2.19.0); additive, no plugin rebuild.
- Plugins can now use one-line text boxes that stay one line. Long or pasted text scrolls sideways instead of making the window grow.
- Pressing Enter in these one-line boxes no longer briefly opens the game chat.
- A new multi-line text box for plugins, with a scrollbar and mouse-wheel scrolling. It can also be read-only, so you can select and copy its text without changing it.
- Bars in plugin lists can now change colour on each row, for example red for a boss and yellow for an elite.
- Eating a potion or food again to add more time no longer makes that buff disappear from the CombatMeter’s Buffs & Debuffs icons, for you and for party members. The buff stays and shows its longer time.
Developer notes
Section titled “Developer notes”- Buff fix:
BuffEffectSyncReadernow treats aBuffEffectof typeRemoveLayer(6)that carries a buff payload (AddBuff/BuffChange) as an UPSERT, not a removal. The game re-applies a stacking consumable on the SAME buff uuid and announces the extended duration as RemoveLayer + BuffChange. The reader used to drop that payload and remove the still-live buff, soICombatLookup.BuffsFor,CombatEvent.BuffChanged(a spuriousRemoved) and everything downstream (CombatMeter’s buff spool and uploads) lost it until the next AOI re-seed. Measured on a real spool: Potion 2033179 was “removed” at the re-eat, then expired by the game at create + 2×600 s.Remove(2)still removes, and a bare RemoveLayer with no payload keeps the old removal behaviour. Pinned byBuffEffectSyncReaderTests. - New
InputElement.SingleLine(init, defaultfalse; existing fields unchanged). A true one-line field that never grows: it uses uGUILineType.SingleLine, becauseMultiLineNewlinemakesInputField.EnforceTextHOverflow()forceWrap. Long text scrolls sideways in uGUI’s draw window, clipped by aRectMask2D. Pasted\r/\nare stripped inonValueChangedand forwarded as a change, never a submit. Enter submits AND ends editing (one press = one submit). The field deactivates while Enter is still held, and the keyboard gate samplesIsFocusedon the throttled tick, so a submit arms a grace:UGuiTextInput.IsFocusedstays true while Enter/KeypadEnter is held plus 0.15 s after release, and the game never opens chat. The grace also covers a submit that hides (parks) the window. Esc and click-away blur report unfocused immediately. Validated in-game. - New
TextAreaElement(Func<string> Get, Action<string> OnChange, float Width = 260f) : HudElement, withLines(init, default 4; the visible line count sets the fixed box height) andReadOnly(init). Enter inserts a newline and keeps focus. There is no submit: give the window its own button that reads the buffer captured fromOnChange(full text, newlines kept). The box is aScrollRect(themed scrollbar, mouse wheel, 8 px / 6 px inner padding) around a field that grows to its wrapped text, so uGUI’s draw window never slices it. A focused caret move or edit scrolls the minimum needed to keep the caret line in view. Caret and selection are clipped to the viewport withCanvasRenderer.EnableRectClipping, becauseRectMask2Ddoesn’t clip InputField’s raw caret renderer.ReadOnlymaps toInputField.readOnly: focus, caret, selection, Ctrl+A and Ctrl+C work; typing, paste, cut and delete are blocked.Getstill replaces the text and may echo throughOnChange, so pass a no-op there. The read-only background is slightly more translucent. When the text fits, the wheel over the box does not reach an enclosingScrollElement. Built; only the no-growth layout is validated in-game so far.new TextAreaElement(() => _draft, s => _draft = s, Width: 320f) { Lines = 6 },new ButtonElement(() => "Import", () => Import(_draft)) - Layout fix: uGUI
InputFieldis itself anILayoutElement(priority 1), and on a priority tieLayoutUtilitytakes the MAX, so a siblingLayoutElementat the default priority could not stop wrapped text from growing the field (height, or width once it overflows). Fixed-box fields (SingleLine, and the outer box ofTextAreaElement) now pinlayoutPriority = 2. The defaultInputElementmode keeps priority 1 and is unchanged. - New
BarElement.FillColorFnandBarElement.LabelColorFn(bothFunc<ColorRgba?>?, init, default null). They are evaluated per refresh, so a pooled or recycled list row can follow the item it currently shows; the staticFill/LabelColorare captured once at build. Returning null falls back toFill, or toLabelColorand then white.FillColorFnis honoured on both theBarStyle.Modernand the Default window bar paths;LabelColorFnonly onModern(the primary left/inside label). - Internal file-size splits, no behaviour change:
WindowBuilder.WindowTokenmoved toWindowBuilder.Token.cs,UGuiTextInput.Buildwas split for the 50-LoC method gate, and the TextArea mode lives inUGuiTextInput.TextArea.cs.
[2.18.0] - 2026-10-05
Section titled “[2.18.0] - 2026-10-05”2.18.0 (minor) — a two-axis pad control for plugin windows. Adds API for plugins (Abstractions 2.18.0); additive, no plugin rebuild.
- A new control for plugins: a square pad you click or drag to set two values at once — for example, where a character looks in Photo Studio.
- Dragging inside a colour picker (the colour square or the hue bar) no longer scrolls the window at the same time.
Developer notes
Section titled “Developer notes”- Drag sink:
UGuiPrimitives.AddDragSinkputs a non-interactableSlider(no target graphic, transition and navigation off) on the raycast target of the XY pad and of the colour picker’s SV square and hue bar, so uGUI drag events stop there instead of bubbling to the window’sScrollRect; the mouse wheel still scrolls the window over them. The pointer itself is still read by the window interaction ticker. - New
XYPadElement(Func<(float X, float Y)> Get, Action<float, float> Set, float Min = -1f, float Max = 1f, Func<bool>? Enabled = null) : HudElement, withSize(init; square side in px, 0 = fill the cell width) andGridLines(init; divisions per axis, default 8; the centre crosshair is always drawn). X grows right, Y grows UP (the top edge isMax). A press or a drag anywhere on the pad sets the value from the pointer, clamped;Setruns on press and on every drag frame, likeSliderElement(no release callback — persist fromSet).Getis poll-diffed, so the dot follows external changes;Enabledfalse dims the pad and ignores the pointer. Window surface only (like the slider). Pointer mapping:Stellar.Application.Services.XYPadMapping(tested); the uGUI builder isWindowBuilder.XYPad.cs(drag via the window interaction ticker, as the colour picker’s SV square).new XYPadElement(() => (_aim.X, _aim.Y), (x, y) => { _aim = new(x, y); SaveAim(); }) { Size = 120f, GridLines = 8 }
[2.17.1] - 2026-10-05
Section titled “[2.17.1] - 2026-10-05”2.17.1 (patch) — stable release of the 2.17 line. Identical code to the 2.17.0 testing build; the number moves only because 2.17.0 was already published to the testing channel and a taken version is never re-minted. No API change; binary-compatible with plugins built against 2.17.0.
- Stable release of the 2.17 line: ReShade support for Photo Studio. With Photo Studio 1.5.1 you can install ReShade from Photo Studio’s page in the Stellar launcher (this needs launcher 2.1.1), pick effect packs inside Photo Studio, and turn effects on and off there. Your photos get the same effects you see on screen.
- Effects that only work at your screen size (many bloom, sharpening, depth-of-field and AcerolaFX effects) are drawn at the photo’s own size in 2x and 4x photos and in other photo shapes, so they look right. Big photos no longer make such an effect switch itself off on screen.
- Portrait, square and wide photos turn off depth-based effects for that photo only, so nothing is drawn in the wrong place. If ReShade is not ready in time, the photo is taken without its effects and Photo Studio tells you so.
Developer notes
Section titled “Developer notes”- Promotion only — no source change besides
FrameworkVersion.Value2.17.1 (and its exact-version test pin). What the 2.17 line adds (IPluginDownloads,IReShade,IReShadeUniforms,CaptureRequest.ApplyReShade,CaptureResult.Notes, isolated capture with bridge 1.1.0) is under the 2.17.0 entry below. - 2.17.0 cannot be re-used for the stable publish:
Stellar-2.17.0.zipis already on the CDN under the bundle-immutability guard and a GitHub pre-release taggedv2.17.0exists, so the stable promotion rides a patch bump.
[2.17.0] - 2026-10-04
Section titled “[2.17.0] - 2026-10-04”2.17.0 (minor) — ReShade support for plugins: effects live and in photos, plus checked downloads into a plugin’s own folder. Adds API for plugins (Abstractions 2.17.0); additive, no plugin rebuild.
- With Photo Studio 1.5.0 or later, you can install ReShade from Photo Studio’s page in the Stellar launcher (this needs the launcher update), pick effect packs inside Photo Studio, and turn effects on and off there. Your photos get the same effects you see on screen.
- Some effects (many bloom, sharpening, depth-of-field and AcerolaFX effects) only work at your screen size. With the updated ReShade bridge, 2× and 4× photos and photos in other shapes are drawn in a separate copy of ReShade at the photo’s own size, so these effects look right (effects that use depth are left out of those photos). With the older bridge, or if that fails, a 2× or 4× photo is taken at 1× and a photo in a different shape leaves those effects out. Photo Studio tells you when this happens. Big photos no longer make such an effect switch itself off on screen.
- Photos in a different shape than your screen (portrait, square, wide) turn off depth-based effects for that photo only, so nothing is drawn in the wrong place; they come back right after.
- A photo takes a brief moment to be prepared for your effects, and waits while ReShade is still loading them. If ReShade is not ready in time, the photo is taken without its effects and you are told so in Photo Studio.
Developer notes
Section titled “Developer notes”- New
IPluginServices.Downloads(IPluginDownloads): https-only downloads pinned by sha256 (verified before anything is written), with a size cap and optional zip extraction filtered byIncludePrefixes, toDownloadRequest.TargetPath(a file, or for a zip a folder that is replaced) inside the calling plugin’s own data folder (DataFolder, shared withIPluginDataStore), which a request can never escape. A malformed request returns “invalid request” instead of throwing. One download per plugin at a time; completes and reports progress on the main thread. - New
IPluginServices.ReShade(IReShade):State(ReShadeState:NotInstalled,Loading,Ready),Enabled,Techniques(ReShadeTechnique: name, effect file, enabled,UsesDepth),SetTechnique(effectFile, name, enabled)(a technique is identified by effect file + name, saved to the current preset),CurrentPreset/SetPreset,SetSearchPathsandChanged. Talks to ReShade through the Stellar ReShade bridge add-on (looked up, never loaded); without ReShade and the add-on it reportsNotInstalledand every member is a no-op. Setters are asynchronous (applied at ReShade’s next frame, visible afterChanged);SetSearchPathsreplaces ReShade’s search paths, persists them and reloads every effect (identical paths are not re-sent).UsesDepthcomes from a scan of the effect source with its includes (ReShade’s standard headers skipped); an effect that cannot be found or read counts as using depth. CaptureRequest.ApplyReShade(init, default true): whenStateisReadyand its effects are on, the capture has ReShade draw its active techniques into the photo’s render target before the readback. Each photo is first prepared for its size (up to 5 s, one ReShade pass per frame; a bridge error ends it at once); alpha is forced opaque after ReShade draws. In a non-screenAspect, enabled techniques that use depth are switched off for that capture only (never saved) and restored afterwards, also on failure.CaptureResult.Notes(init, default empty): player-readable remarks on a successful capture, e.g. “ReShade was not ready — photo taken without it.”ReShadeTechnique.SizeLocked(init, non-positional): the effect declares a texture sized from the screen (BUFFER_WIDTH/BUFFER_HEIGHT,BUFFER_SCREEN_SIZE,ReShade::ScreenSize, … directly, through a#define/const, or through an include;: COLOR/: DEPTHtextures excluded). ReShade 6.8.0 keeps named textures in one global list, so such a texture created by the screen-size permutation is reused by a larger photo’s permutation, whose passes then read its top-left 1/scale region magnified (runtime.cpp:2035-2084). Unknown counts as locked. While an active technique is size-locked, a window-shaped 2×/4× photo is taken at 1× with ReShade (Width/Heightare the real 1× size) and a shaped photo switches the size-locked techniques off for that photo; notes: “Some ReShade effects only work at screen size, so this photo was taken at 1×.” / “Some ReShade effects only work at screen size, so they were left out of this photo.” No size-locked effect is ever compiled at a non-screen size, so a large-size compile failure (AcerolaFX AutoExposurenumthreads(3600,1,1)at 4×) can no longer disable it on screen (runtime.cpp:2164, 3763-3775).- Isolated capture (StellarReShadeBridge 1.1.0+, detected per export with
GetProcAddress;rsb_version()stays 1): when the photo is not screen-sized and an active technique isSizeLocked, the photo is drawn in a separate effect runtime sized to the photo (rsb_isolated_*), with everyUsesDepthtechnique requested off (that runtime has no game depth). Config:<game>/stellar/reshade-capture/StellarPhoto.ini(+StellarPhoto.ini.preset.ini, written by the bridge). Bounded at 20 s; any failure (missing exports, a refused begin, a negative state, state 4, a render error, the timeout) ends the session and takes the 1×/skip guard above with its note; the session is ended on every exit path (plan, throw, host loss). New note on a window-shaped isolated photo that left depth techniques out: “ReShade effects that use depth were left out of this photo.” Warm-up: a fresh isolated runtime has no history, so adaptive effects start from zero (prod80 Bloom’s previous-frame average luma turned a 4× “Stella Medium” photo almost white). When an active effect is temporal — aframetime/timer/framecountuniform or aPrev/Last/History/Accum/Adapt-named texture, unknown = temporal; 20 of the 67 size-locked effects in the measured packs — the isolated runtime renders once per frame for ≥ 2 s and ≥ 30 frames before the real render, inside the 20 s bound. The camera is rendered once; every render starts from a pristine copy (the bridge writes the result back into the queued texture), and both targets are released on every exit path. Diagnostics:[PhotoCapture] reshade isolated=1 … outcome= state= renderCode= framesWaited= warmups= … addon=; the game-runtime line now readsisolated=0. - New
IPluginServices.ReShadeUniforms(IReShadeUniforms):bool SetUniformOverride(string? effectFile, string variable, string? value)andvoid ClearUniformOverrides(). Holds a uniform at a value in the game’s ReShade runtime and in the isolated capture runtime, re-applied by the bridge after every effect reload, preset switch and runtime re-creation (bridge 1.1.0+rsb_set_uniform_override/rsb_clear_uniform_overrides, bound optionally viaGetProcAddress). Null/empty effect = every effect; values are 1–16 comma-separated numbers ortrue/false(validated before sending); a null/empty value removes. False for bad input or a 1.0.0 bridge; overrides set before the bridge binds are held and sent on bind. Per plugin:ClearUniformOverridesremoves only that plugin’s overrides, and all of them are removed on plugin unload. (IReShadeis at the 8-member cap, hence the separate interface.) First use: AcerolaFX’s_MaskUI, which discards its output in this game. - Capture warm-up: waits (draws nothing) while the bridge reports ReShade loading, and a frame’s drawn count only counts when no load followed it — ReShade’s preset re-apply after the first draw at a new size can switch on further techniques whose compile then starts (runtime.cpp:3722), and a render during it draws 0 (runtime.cpp:3810). A real render that drew nothing (no hard error) is discarded and warmed up again. Deadline: 5 s not counting loading, 20 s at most. Each failure has its own note: “ReShade was not ready — photo taken without it.” (timeout), “ReShade could not draw into the photo — photo taken without it.” (bridge error), “ReShade drew nothing into the photo in time — photo taken without it.” (real render kept drawing nothing), “None of the active ReShade effects work in this photo shape — photo taken without them.” (every active technique had to be switched off — before, this waited out the 5 s).
- The effect scans no longer follow an
#includewritten inside a/* */comment (SweetFX FXAA.fxh documents one), which had made FXAA count as using depth.
[2.16.0] - 2026-10-03
Section titled “[2.16.0] - 2026-10-03”2.16.0 (minor) — Photo mode: hide yourself and hide effects by whose they are. Adds API for plugins (Abstractions 2.16.0); additive, no plugin rebuild.
- Photos can now hide your own character — together with your pet and your Battle Imagine — on your screen only. You can still move and fight, and other players still see you.
- Photos can hide skill, buff and hit effects by who caused them: yours (including your pet’s and Battle Imagine’s), your party’s, other players’ or monsters’ (including boss warning areas). Scenery like waterfalls and lamps is never hidden, and everything comes back when you turn the switch off.
Developer notes
Section titled “Developer notes”VisibilityLayersgainsSelf(32),EffectsMine(64),EffectsParty(128),EffectsOthers(256),EffectsMonsters(512) andVisibilityLayerSets.Effects; all go throughISceneVisibility.Hidelike the existing layers. Values of the old layers are unchanged.- Self uses the game’s own photo entity-show counters (Oneself + SelfPet). Effects are classified by the effect’s caster (
FromUuid, elseBelongUuid); a summon resolves to its owner viaEntitySummonAppeared, else the entity’s ownTopSummonUuid/SummonUuid; an effect with no resolvable owner is never hidden. Only effects that were visible when hidden are ever shown again; ZEffect pooling is guarded by a uid read-back. - Effect creation hooks (
ZEffectManager.AddEffectDisplay,ZEffect.Init) install lazily on the first effect-layer hide. A failed show-back is retried on the next re-assert (IVisibilityBackend.HasPendingRestore). - With
STELLAR_DIAGNOSTICSon:[EffectHide] apply/classified/unresolved/zero-caster/born-hidden/owner-lookupand[PhotoVis] holds after Selflines (all capped).
[2.15.0] - 2026-10-03
Section titled “[2.15.0] - 2026-10-03”2.15.0 (minor) — Photo mode support for plugins: screenshots and looks, a free camera, posing, a scene freeze, photo shapes and lights. Ships the never-released 2.13.0 and 2.14.0 (the release tooling folds their notes in). Adds API for plugins (Abstractions 2.15.0); additive, no plugin rebuild.
- Free camera posing: pick a person — you, another player or an NPC — and set their pose, the exact moment of it, a facial expression that stays, where their head and eyes look, and which way they face. Other players and NPCs are posed as a copy only you can see. Posing works with the free camera off too, and leaving the free camera keeps everyone as you posed them — they return to normal when you change zone, a cutscene starts or you disconnect.
- The free camera orbits the copy you are posing, and freezing the scene freezes posed copies too.
- Freezing the scene now pauses the whole world, your own character included: skills stop mid-cast, effects and animations hold still, and players and monsters stay where they were, in the pose they had. While frozen your character ignores movement and skill keys (the camera still turns). The free camera, the Photo Studio panel, screenshots and all your plugin windows and hotkeys keep working while paused, and leaving the free camera keeps the world paused until you unfreeze. The game itself keeps running on the server: a fight goes on, and damage taken meanwhile shows when you unfreeze. If something hits you while frozen, your own character may still flinch in place.
- Picking someone who is already doing an emote shows that emote and how far along it is, and pausing holds it right where it is.
- Photo lights: place lamps around the people in your shot — each with its own colour, strength and reach — that light the ground and scene, let characters take their light, and give anyone a key light from one side and a coloured rim. Lights stay when you leave the free camera and are gone, with everything back as it was, when you change zone, a cutscene starts or you disconnect.
- Screenshots can now be taken in other shapes — portrait, square or wide — at full detail: the camera draws the photo in that shape instead of cropping a wide shot, and very large sizes are reduced to what your graphics card allows.
Changed
Section titled “Changed”- In the free camera you can now click NPCs to orbit them, not just players.
Developer notes
Section titled “Developer notes”- Release note: the launcher manifest for 2.15.0 shows three player bullets twice — the 2.13.0/2.14.0 bullets were folded into this section by hand AND by
changelog.py, which already folds never-published versions. Published manifests are immutable; this section is corrected for the record. ColorPickerElement.ShowAlpha(init, default true): false hides the Opacity slider for colours with no alpha (Photo Studio lights). Additive — the two-argument constructor is unchanged.- New
IPluginServices.Lights(ILights:IsAvailable,AddLamp,UpdateLamp,RemoveLamp,PeopleLevel,SetPersonLight,ResetAll,Released) and domain typesLampId,LampSettings,KeyLight,RimLight,PersonLight,LightLimits(spec devkit 2026-10-03-photo-studio-lights; reconfree-camera-recon.md§ Run 13). Lamps = a GameObject with a Unity pointLight+ the game’sBokura.Rendering.MultiLight(lightLayer = Everything,type = Common,maxDistance = 128), set up inactive then activated soMultiLight.OnEnableadds it to the light cluster; off = inactive; removal = inactive +Destroy. The cap (8 per plugin) is enforced framework-side. The character gate isCameraManager.weatherParamsVolume_.creaturePointlightColorIntensity: raised = snapshot allparameterListoverride flags (101 on release_3.7) + the parameter’s value +active, every flag off, only the gate overriding,active = true; restored (value, flags, active) exactly when no lamp is on or every level is 0 (LightGate,LightGatePolicy). Key light =RenderCompBase.SetFixedLight(ref ModelFixedLightData{LightParms}, 0x3FFFFFFF)→_CameraLightParm(a camera-space direction, w = 1); rim =SetFresnelEffect(1, colour × strength, (−0.7, 1, 1, 1), All); both written back from a per-material snapshot taken before the first write (ModelLightSnapshot; the game’s own Fresnel “off” leaves colour + params). Plain reflected calls only —ModelFixedLightDatais by-ref and is never HarmonyX-patched. The posed copy / NPC stand-in comes from the posing backend; lit people are re-resolved on everyIPosing.Changed.FreeCameraReleaserreleases lights after the camera and before posing (people written back while their copy lives → lamps → gate);ReleaseAllis idempotent and raisesReleasedonly when it ended something; the per-plugin facade ends a plugin’s lights on unload. Diagnostics (STELLAR_DIAGNOSTICS=1):[Lights] lamp add|remove,[Lights] gate raised|restored …,[Lights] person <uuid> key|rim written|restored. Hardening (lights review 2026-10-03): every key / rim apply first reads any material the model gained since (equipment) so it is written back too; a posed copy cloned from a lit person saves that person’s originals for the values we wrote; a gate write that throws is rolled back from the snapshot and the restore attempts every write; the gate is put back only while the volume still holds exactly what we wrote — if the game wrote it since (a cutscene, its own weather) it is left to the game — and a volume the game already has active keeps its own overrides (only ours is added); lamps live on a persistent root (DontDestroyOnLoad), never in a streamed scene chunk. - New
ICameraOverride.TryProjectToScreen(Position3D, out ScreenPoint)+ domainScreenPoint(X, Y, InFront): where a world point appears through the camera rendering now (the free camera while held, else the game camera), in screen pixels top-left (theIEntityPickerspace) — for on-screen markers (Photo Studio’s lamp markers). A plugin that IMPLEMENTSICameraOverridemust add it. WindowSpec.Passive(Borderless windows): a visual-only layer with no content padding and no click blocker, so presses reach the game and every window behind it — for full-screen guides/overlays (Photo Studio’s frame guide).IScreenCapture.PlanSizeis a new interface member: a plugin that IMPLEMENTSIScreenCapturemust add it.- Photo shapes (spec 2026-10-03 photo-studio-portrait-capture): new
CaptureRequest.Aspect(CaptureAspect?, a width:height ratio of whole numbers, valid 1:4 … 4:1; null = the window’s shape, unchanged) andIScreenCapture.PlanSize(CaptureRequest)(the real output size before the shot — scale caps, shape and the GPU’sSystemInfo.maxTextureSizeincluded; empty for an invalid scale/shape). Pure maths on the newCaptureSizing(OutputSize,EffectiveScale,GuideRect,VerticalFieldOfView,MaxLongSide,MaxPixels) withCaptureSize/NormalizedRect: a shape’s long side = the window’s long side ×Scale, the short side follows the shape rounded to an even pixel count (2× on 1920 × 1080: 9:16 = 2160 × 3840, 4:5 = 3072 × 3840, 2:3 = 2560 × 3840, 1:1 = 3840 × 3840, 21:9 = 3840 × 1646); over the long-side limit (16384 or the GPU’s, whichever is smaller) or 64 MP both sides shrink by the same factor (a window-shaped capture keeps its 4× → 2× → 1× halving). The camera renders straight into a target of that size withCamera.aspectset to it for that one render (and, for a shape wider than the window,fieldOfViewnarrowed so the horizontal angle is kept) and both put back before anything else renders (CaptureLensOverride: automatic aspect again when it was the window’s) — so the photo frames exactlyGuideRect, the largest centred rectangle of the shape. Property writes only, no new hook. An invalid shape fails with a readable error; a failed shaped 4× grab retries once at the 2× shape. A PHYSICAL camera (usePhysicalProperties, which the game sets) stays physical for the shaped render: the framing rule is applied through an explicit VerticalgateFit+ the matchingfocalLength(neverfieldOfView, which would move the focal length), computed from the on-screen gate fit (Vertical/Horizontal/Fill/Overscan/None) by the pureCaptureLensPlanner; every property it changed is put back exactly once, even when a write or a restore step throws. Diagnostics log the lens set/restore and whether the next frame’s whole lens equals the pre-shot one ([PhotoCapture] lens next-frame ok=…). NewIScreenCapture.MaxTextureSize(the GPU limitPlanSizeapplies — pass it toCaptureSizing.EffectiveScale). - New
IPluginServices.Posing(IPosing:IsAvailable,NearbyPeople,Expressions,Select,ResetAll,Changed;IPoseTarget:State,PlayAction,Moment,SetExpression,SetLook,Aim,Yaw,Reset) and domain typesPersonKind,PersonInfo,ExpressionInfo,LookMode,LookPart,PoseTargetState,PoseResult.IPosing.IsAvailable=IClientState.IsWorldActiveand outside the posing scene-change settle window (~2 s after a leave/enter) — no camera override needed; a camera release resets nobody. The scene end (FreeCameraReleaser.Release:Game.OnLeaveSceneprefix /SceneChanged, cutscene, game photo mode, logout, framework unload — the reasons the scene freeze already ends on) resets every touched person in the same call, aZEntityMgr.RemoveEntityprefix releases a despawning person, the per-plugin facade releases a plugin’s people on unload, and a scene-freeze end re-applies held pauses. - Game calls (devkit
docs/recon/photo-posing-recon.mdruns 4–5): you = live model,ZAnimActionPlayMgr.PlayAction(id, syncServer: false, …)/SetActionPersistTime/ResetActionandLuaAsyncBridge.SetEntityRotation(an action start and a reset each ride one NewMove — the game’s own photo behaviour); players =CloneModelForPhoto+CameraFrameCtrl.SetTargetEntityVisible(e, false)…RecyclePhotoModel; NPCs =NpcTable.ModelID→LuaAsyncBridge.GenerateNormalModelAsyncByLua(…, isCameraScene: true)withApplyModelBaseIdleByLua, removed byZModelManager.RecycleModelByLua(the live NPC is never posed); held expression =ZModel.SetLuaAttrEmoteInfo(faceId, -1, true); head/eyes =ZModelHelper.SetLookAtIKParam/SetLookAtTransform/SetLookAtPos/LuaWorldPosToLocal. Copies and NPC models send nothing. IPosing.TryGetVisiblePosition(the copy / stand-in position for an orbit centre, allocation-free); copies are capped at the game’sPhotographTeamMemberLimit(PC[1], 30 in release_3.7: you + 29 players; NPC models 30) withPoseTargetState.Full/PoseResult.Fullbeyond it; whileISceneFreezeis frozen, posed copies/models get the model-level freeze stage (AnimComp.Speed0, prior restored on unfreeze and before release).ISceneFreezeis a global time pause (Photo Studio scene-stays spec amendment 2026-10-02 late; devkit reconfree-camera-recon.md§ Run 9):GameClockPausesavesTime.timeScaleand sets 0; a HarmonyX prefix onUnityEngine.Time.set_timeScale(float)(TimeScalePatch, by-value float — allowed byIl2CppPatchSafety) holds the game’s own writes while paused (hit-stopZTimeScaleShowInfo.OnStopwrites 1.0) and keeps the latest as the restore value (else the saved one; never ≤ 0). A watchdog (TimePauseWatchdog) at the framework’s global rate, OUTSIDE the world gate beside the login / loading-screen probes (never on the world-gatedIFramework.Update), re-asserts 0 if the clock was set behind the hook, resumes a pause its freeze already dropped and releases a pause outside the world; a stall check after every paused frame’s tick (no framework tick for 10 s, persisting over 3 checks and 1 s — one long hitch is no stall) and the paused driver’s destruction release too. Every such release is the FULL release path (FreeCameraReleaser: camera, posing, every freeze token → animation replay + gate disarmed, hold stopped, clock resumed, input shield), never a bare clock resume; framework unload resumes as well. While paused the framework tick runs from real time (StellarPausedTicker, enabled only while paused;TickPacerkeeps one dt and one tick per frame across both drivers — theInvokeRepeatingpath is unchanged), and the free camera’s acquire / release cut (CinemachineBrain.m_DefaultBlend= Cut for the pause, the game’s blend put back on resume). Kept from the per-entity freeze: theModelGoCompposition + rotation hold for remote movers (never the local player or their own mount) and the deferred removal of a monster killed while frozen. New (owner report on the TEST window, recon runs 11–12): at timeScale 0 the ECS animator’s clock is frozen (0 animation events over a 20 s pause) but the game keeps REQUESTING states for movers (a walking monster: 9PlayBaseStatein 20 s), and one request re-poses a model with the clock stopped (45 % of its region changed). So a run-original gate onECSAnimController.Play{Base,Upper,Additive}State/PlayManualClipand their internalplay…twins holds the requests of the press’s entities’ controllers (never the local player, their mount or a posing copy) and replays the latest per layer on unfreeze, each only while its entity still serves the same (pooled) controller (AnimRequestGate; kill switchSTELLAR_FREEZE_ANIM_GATE=0). The gate is four TYPED HarmonyX prefixes (AnimGatePatch, by-position arguments, exact-signature check) on the internalplayBaseState/playUpperState/playAdditiveState/playManualCliponly — the public, sequence and state-end requests all funnel through them (release_3.7 ISIL) — so an unfrozen call costs a counter and a field check and allocates nothing; installed on the first freeze for the session. A press that does not know the local player tracks nobody. The local player’s movement / combat input is masked while paused (ZIgnoreMgr.SetInputIgnore, maskShieldMask.PauseBits=0x10FFB6679, camera bits open) as a second layer of the free camera’sEGmsource:InputShieldServiceowns that one source and moves it between unions by the difference only (the game counts each bit per source). Removed: the attr factor freeze, theAnimCompBase.set_Speedgate, the ECS layer gate and its native play detours (EcsPlayDetours), the effect freeze atAddEffectDisplay/ZEffect.Init, the appear / ride-up re-checks and the combat-freeze diagnostics capture; diagnostics keep one freeze on/off summary (time scale saved/restored, game writes held, drifts, positions held, animation requests held / replayed / stale, the gated entry points’ calls this freeze and since install, deferred removals).- While frozen, Unity’s scaled time stops for every plugin, not just the game:
Time.deltaTimeis 0 andTime.timestands still,WaitForSeconds,UniTask.Delay(defaultDelayType.DeltaTime) and anything else on scaled time wait until the unfreeze. UseTime.unscaledDeltaTime/Time.realtimeSinceStartup,WaitForSecondsRealtimeorDelayType.Realtime/IFrameworktimers for anything that must keep running. The position hold covers players and monsters within 80 m of you (FreezeKinds.HoldRadius); farther ones may glide in their paused pose.ISceneFreeze.Changedhandlers each run on their own (a throwing one is logged once); one throwingIFramework.Updatesubscriber no longer stops the others. - Known limits: the server keeps driving the local player’s hit reactions while frozen, so your own character can still
re-pose in stop-motion when something hits you (its requests are not held — holding them would also hold your own
movement); a disconnect ends the freeze only through the game’s logout (
IClientState.Logout). IPosing.TryGetCurrentAction(person, out actionId, out moment): the running action (posed copy/model first, else the live entity’s model) via compiledZModel.GetLuaAttrActionInfo{ActionId,TotalTime,PassedTime}reads, liveness- and settle-gated, 0 sends; settingIPoseTarget.Moment0–1 before anyPlayActionadopts and holds that action without re-playing it (a reset only releases the hold on an adopted action). The read reports the posed model’s own state whenever one exists; an adopted action the opened model lacks is dropped. Diagnostics:[Posing] action-raw(≤ 5/s).IEntityPickeralso returns NPCs.HarmonyGameMethodHookerskips abstract overloads (they cannot be patched).- Diagnostics (
STELLAR_DIAGNOSTICS=1):[Posing] open …/[Posing] npc model … loaded=…lines, a[Posing] despawn prefix fired on managed thread …line on everyZEntityMgr.RemoveEntityprefix, and[Posing.Send] svc=… method=…for every outgoing RPC while a person is selected. - Whole-feature review:
Expressionsand the photo-member limit read are each behind a bounded retry latch (NegativeProbeCache, same pattern asGameVisibilityBackend’s probes) — an empty/zero answer or a throw is cached and retried at most every 5 s instead of re-running the Lua query (and re-warning) on every 10 Hz panel poll. - HarmonyX patch safety (unreleased-code fixes, review of the freeze-crash fix): the hooker now refuses — logged as an
error, never patched — any method with a by-ref value type other than an integer primitive (blittable struct, enum,
float, double, char) or a register-sized / unknown-size IL2CPP struct RETURN, besides the by-value 1/2/4/8-byte and
by-ref IL2CPP struct parameters it already refused. Measured on the game’s CoreCLR 6.0.7 (TEST): Il2CppInterop’s
trampoline hands a by-ref
Vector3/Quaternion/QualityDatato the original with only its first 8 bytes and writes 8 garbage bytes back over the caller’s; aref floatmakes it throw so the original never runs. By-value blittable structs were byte-identical.QualityGradeSetting.ApplyAllData(ref QualityData)(the render-quality re-assert signal) is therefore a native detour (QualityApplyDetour, original first, then the signal), the diagnostics-onlyMoveComp.MoveGo/MoveGoByCurve/MoveGoBySpeedprefixes are gone, and every native detour (QualityApplyDetour) goes live only on its full exact interop signature (NativeSignature), else an error line.
[2.14.0] - 2026-10-01 (never released — shipped in 2.15.0)
Section titled “[2.14.0] - 2026-10-01 (never released — shipped in 2.15.0)”2.14.0 (minor) — Free camera support for plugins. Adds API for plugins (Abstractions 2.14.0); additive, no plugin rebuild.
- Free camera support for plugins: move the camera freely around your character, freeze the moment (characters, NPCs, pets and effects stop), and strike a pose with your own emotes.
- A hotkey you set yourself can no longer end up sharing its key with another action’s default.
Developer notes
Section titled “Developer notes”- New
IPluginServicesmembers:CameraOverride(exclusive Cinemachine vcam takeover; 60 m hard cap from the local player; framework release on zone change / cutscene / game camera mode / disconnect / unload / frame-handler exception withCameraReleaseReason;LookAtCamera()snapshot-restore),InputShield(ref-countedZIgnoreMgrmask underEIgnoreMaskSource.EGmplus the text-fieldKeyboardInputGatefor every game key, raw key/mouse reads),SceneFreeze(everyEntityDictentity: effects viaSetEffectFreeze+AddEffectDisplaypostfix, animation viaSkillStageTimeFactorthenAnimComp.Speedwhere the drawn speed stays above 0, drawn-position hold of movable kinds within 80 m in LateUpdate with a 0.3 ms/frame self-check and a release snap, entities appearing while frozen via anAddEntitypostfix),Emotes(the emote wheel’s unlocked list and check-then-play path through the Lua VM),CombatState(attr 104 +SetLocalCombatData/SetLocalAttrInBattleShowpostfixes, no polling),EntityPicker(chest projection). - Kill switches:
STELLAR_FREECAM_OFF=1,STELLAR_FREEZE_NO_POSITIONS=1. IHotkeys.MigrateSavedBinding(actionId, from, to)(SavedBindingMigration): one-time move of a player-saved chord. Resolution now lets a saved binding beat a suggested default whatever the declare order (it used to skip the collision check); default-vs-default is unchanged.- A per-frame
FreeCameraFrameDriverMonoBehaviour exists but is disabled unless a camera override or a position hold is live. IInputShieldHandle.TextFieldFocused: true while a Stellar overlay text field has keyboard focus (the same source as the keyboard gate), so a free-camera consumer can ignore movement/edge keys typed into a panel search box. False once disposed.
[2.13.0] - 2026-09-30 (never released — shipped in 2.15.0)
Section titled “[2.13.0] - 2026-09-30 (never released — shipped in 2.15.0)”2.13.0 (minor) — Screenshot support for plugins. Adds API for plugins (Abstractions 2.13.0); additive, no plugin rebuild.
- Screenshot support for plugins: hide the interface, take high-resolution screenshots and apply camera looks.
- Plugins can now sharpen the picture for screenshots (supersampling, sharper shadows) and set the time of day; your own graphics settings come back exactly when they stop.
Developer notes
Section titled “Developer notes”- Four new services on
IPluginServices:ScreenCapture(IScreenCapture— PNG/JPG at 1×/2×/4×, long side capped at 16384 px, a 4× failure retries once at 2×, never throws),SceneVisibility(ISceneVisibility— reference-counted hiding of the game HUD, the Stellar overlay, nameplates and other players; a plugin’s tokens andChangedhandlers are released when it unloads),RenderLook(IRenderLook— one active look framework-wide;PlayModestrips depth of field and film grain) andPhotoMode(IPhotoModeState— the game’s camera / selfie mode and cutscene state). - Looks: the framework owns one global
Volume(priority 10000, runtimeVolumeProfile) and writes the game’s ownZRenderPipelinecomponents into it —ZDofVolume(Bokeh),ZColorAdjustmentVolume,ZWhiteBalanceVolume,ZColorLookupVolume(256×16 / 1024×32 PNG strips),ZBloomVolume(intensity_UE/threshold_UE),ZUnityVignetteVolume,ZFilmGrainVolume(every parameter written, as the game’s cutscene grain track does:filmTypeCustom + a framework-generated 256×256 linear mid-grey noise texture in R=G=B=A, Tiling (1,1), white Color — the bundled Thin1 texture showed no visible grain). It never edits the game’s profiles; omitted groups drop their overrides and revert to the game’s value. Evidence: devkitdocs/recon/photo-studio-render-recon.md(DXVK section). - Capture renders the main camera into an N×
RenderTextureat end of frame (clean: no game UI, nameplates or overlay) and reads it back as RGBA32; PNG encoding is a managed encoder run off the main thread; the result completes on the main thread. - Hide switches:
ZUiRoot.SetUIInvisible,HudMgr.SetHudSwitch(…, ECamera),CameraFrameCtrl.SetEntityShow(per-type hold counters shared with the game’s camera panel:11OtherPlayer hides everyone; keep-party hides Stranger6+ Chum2+ Union4and never writes11, because 11 is a master switch checked before Team; one show per hide issued;ZEntityMgr.ForceRefreshCharVisible(ETakePhotos)on a party change while keep-party is held), and the framework’s window + layout-edit canvases (the toast canvas stays visible). Photo mode:CameraFrameCtrl.Record/ResetCameraInitialParameters+CameraStateSelfPhoto.OnEnter/OnExit; cutscene:CutsceneManager.Play/afterStop, seeded fromInCutscene. Every game member is resolved by name at runtime and fails open with one warning. - Two more services on
IPluginServices:RenderQuality(IRenderQuality—Request(RenderQualityRequest)returns a token; reference-counted union across plugins:Supersample= render scale 2.0 + TAA on,HighShadows= 4096 shadow map, 3 cascades, soft shadows;Liveread-back,Capabilitiesper lever) andTimeOfDay(ITimeOfDay—Pin(hour)returns anITimePin; newest pin wins, disposing falls back, the last dispose hands time back to the server;SetHouron a held pin). A plugin’s tokens and pins are released when it unloads. - Each lever captures the game’s value before the first write and restores it when the last token goes; a value the game writes while a lever is held (its grade re-apply, the settings panel) becomes the restore target. Every write happens only when the live value differs. Levers:
ZRenderPipeline.asset.renderScale;QualityGradeSetting.EnableAA+applyEnableAA;ZShadowCastPassshadowSettings.shadowmapResolution/cascadesCount/supportsSoftShadows; time of day throughPanda.LuaAsyncBridge(SetWeatherIsUpdateFromServer,SetCurWeatherTime,GetCurWeatherTime24,GetWeatherIsUpdateFromServer) —ZServerTimeis never touched. - Re-assert is event-driven (no timer): scene/phase change, postfixes on the game’s quality-apply entry points (
QualityGradeSetting.ApplyAllData,applyRenderScale,applyShadowGrade, the resolution/grade setters and the settings panel’s Lua wrap),ZShadowCastPass.OnInitialize, the game’s ownLuaAsyncBridgetime calls (bridge + Lua wrap), and the end of the game’s photo mode or a cutscene; it drains on the next tick onceIsWorldActive. Hook list and known gaps: devkitdocs/recon/photo-studio-render-recon.md§ Render quality + time of day recon. - Capture render-scale guard (drops a supersampled render scale to 1.0 for the grab frame) ships OFF;
STELLAR_CAPTURE_SCALE_GUARD=1turns it on for the in-game measurement.HarmonyGameMethodHooker.PostfixStaticOverloadsadded for static game methods. The render-quality / time-of-day game hooks install lazily on the firstRequest/Pin(never at boot);STELLAR_RQ_NO_APPLYALLDATA=1skips just the by-refApplyAllDatahook.
[2.12.0] - 2026-09-30
Section titled “[2.12.0] - 2026-09-30”2.12.0 (minor) — Showing and hiding the Combat Meter no longer freezes the game, the meter opens much faster, and layout edit mode no longer lags. Adds a diagnostics API for plugins (Abstractions 2.12.0); binary-compatible, no plugin rebuild.
- Hiding and showing the Combat Meter (or any mod window) no longer freezes the game for a moment. Toggling it quickly is now instant.
- The Combat Meter appears much faster when it first opens, without the short freeze it used to cause.
- Layout edit mode no longer makes the game lag while it is open.
Developer notes
Section titled “Developer notes”- Measured on the owner’s client with the new
[Hitch]recorder: every meter show was a 350-476 ms frame (mount:combatmeter.main, 9/9 toggles, no GC) and CooldownBar 57-72 ms. After: the meter’s first build is 14.7 ms for its 20 raid rows (29 ms first paint), and hide/show triggers no rebuild. WindowRendererimplementsIWindowParking:SetVisible(false)deactivates the built tree instead of destroying it; a re-show reactivates it on top of its stacking tier.Remove()still destroys.ConditionalElementbranch content is built the first time its branch becomes active (the active branch still builds inside the first paint). The CombatMeter main window previously built its list view and both party layouts (49 meter rows) to show one of them.- Meter-row debuff cells are created the first time a slot is visible (default 4 of the 12-cell pool; none while the block is hidden), with their click wiring.
IWindowControl.SetRectskips the persist when the rect equals the saved one: the layout toolbar’s per-tick re-centre was rewriting the whole framework config ~28x/s.- New
HitchProbe(Abstractions, diagnostics-gated viaSTELLAR_DIAGNOSTICSor aHITCHflag): a rendered frame over 60 ms logs a[Hitch]line with the framework tick and its segments, each plugin’s Update, window mount/apply/destroy and build phases, font-atlas refreshes, config saves, hook time, and managed + IL2CPP GC counts.
[2.11.0] - 2026-09-26
Section titled “[2.11.0] - 2026-09-26”2.11.0 (minor) — Mods can now tell which spec nearby players are playing, the moment they come near you, and know the buffs players already have when they arrive. Adds API for plugins (Abstractions 2.11.0).
- Mods can now see which spec a nearby player is playing as soon as they come near you, even in town before any fight. Combat Meter uses this to show specs straight away.
- Mods now know the buffs a player already has the moment they come near you, instead of only after those buffs change.
Developer notes
Section titled “Developer notes”- Spec from talent root buffs: each of the 18 specs has exactly one root talent buff (“<Spec> Spec”); the map is derived at runtime from TalentStage → TalentTree → TalentTable effects (18 pinned fallback constants). Buff-derived spec is authoritative over cast inference; the swap gap holds the previous spec (expires after 10 s); a snapshot with no root clears it; Battle Imagine transform ids are not class changes. Evidence: devkit
docs/recon/spec-from-public-wire-data.md(2,111/2,111 snapshots; owner-confirmed live). - New API (additive):
ICombatSpec.TryGetTalentSpec(EntityId, out int);CombatEvent.SpecChanged(TimestampMs, TargetId, OldSubProfessionId, NewSubProfessionId, FromTalent)(once per real change, never for the swap gap);CombatEvent.EntityBuffsSeeded(TimestampMs, TargetId, Buffs)—SyncNearEntities/ EnterScene now decodeEntity.buff_infosand replace the entity’s buff set silently, one seed event per entity (live deltas unchanged; a later delta for a seeded buff arrives as Refreshed, its expiry as Removed).BuffsFor/LocalBuffsreturn cached read-only snapshots. STELLAR_WIRECAPdebug capture is now lossless for large frames (node budget 4096 → 262144, hex for short opaque byte fields, raw base64 when a frame still truncates).
[2.10.0] - 2026-09-26
Section titled “[2.10.0] - 2026-09-26”2.10.0 (minor) — Lets plugins save the loadout you’re wearing into another loadout slot through the game’s own Save, and tell whether you have unsaved loadout changes. Powers Loadout Switcher’s new copy button. Adds API for plugins (Abstractions 2.10.0).
- Plugins can now copy the loadout you’re wearing into another of your loadout slots, using the game’s own Save, so the game confirms it with its usual “Loadout saved successfully!” message.
Developer notes
Section titled “Developer notes”- New
ILoadoutSaveonIPluginServices(LoadoutSave):SaveCurrentToAsync(index)runsweapon.AsyncSaveRolePlan(index)on the main-thread drain (refuses without dispatch when the target is the worn plan, unknown, or a switch/save is in flight; 8 s timeout; re-arms the SyncProjectList refresh on success).HasUnsavedChanges= the game’sCheckRolePlanIsChange(), carried as anUNSAVEDrow on the existing LiveStateChunk/RefreshChunk (event-driven, no timer). Measured on the owner’s client:SaveProjectaccepts a non-current plan id and copies the LIVE setup (class, gear, modules, skills, talents, Battle Imagines); the plan name is not copied.
[2.9.0] - 2026-09-25
Section titled “[2.9.0] - 2026-09-25”2.9.0 (minor) — The party meter gains shields, buffs & debuffs, and voice controls; plus a dungeon-teleport archive fix and a band-instrument tone relay. Adds render/data API for plugins (Abstractions 2.9.0).
- The party meter can now show each member’s shield (a band over their HP bar) and their active buffs and debuffs as icons.
- You can click the voice icon on your own meter row to cycle your team-voice mode, and open a party-voice menu from the meter.
- A dungeon fight is no longer split into two separate records when you get teleported within the same instance mid-fight.
Developer notes
Section titled “Developer notes”- MeterRowData status-effect channel (DebuffSlot + IsBuff, Debuff0..11, DebuffColumns, DebuffOverflow, ShowDebuffs, DebuffCellSize, BarShieldFraction, OnDebuffClick); WindowBuilder renders the 2-row colour-coded block (variable columns, per-cell click). IGameDataCombat.AllBuffs() enumerates the buff table; GameTextureElement.Fill stretches an icon to its box. Shield via AttrShieldList → EntityVitals → HP-spine band; InstrumentToneRelayPatch relays band tone listener-side.
[2.8.5] - 2026-09-23
Section titled “[2.8.5] - 2026-09-23”2.8.5 (patch) — Settings → Hotkeys now lists numbered actions in natural order, so “Apply Loadout 10” sits after 9 instead of between 1 and 2. Cosmetic; binary-compatible, no plugin rebuild.
- In Settings → Hotkeys, an action numbered 10 or higher now appears in the right place in the list (after 9), instead of being sorted as if it came right after 1.
Developer notes
Section titled “Developer notes”FrameworkVersion.Value→ 2.8.5.HotkeysPanelsorts within a plugin group byNaturalOrder.Compare(numeric-aware) instead of an ordinalstring.Compareon the action id — digit runs compare as integers, soloadout.apply.10sorts afterloadout.apply.9. Bindings are unaffected (keyed by id, unchanged). +8NaturalOrderTests.
[2.8.4] - 2026-09-22
Section titled “[2.8.4] - 2026-09-22”2.8.4 (patch) — Loadout Switcher Deep-Slumber fixes: switching builds now rebuilds the whole Psychoscope tree reliably, and shared factors are no longer taken from your other build unless there’s truly no other way. Application-only; binary-compatible with every existing plugin, no plugin rebuild.
- Switching between builds now rebuilds the whole Psychoscope tree properly instead of sometimes leaving part of it locked.
- When you switch to a build, a Phantom Factor it shares with another build is no longer pulled out of that other build if you already have a spare, or if the build you’re switching to already has it in place.
Developer notes
Section titled “Developer notes”FrameworkVersion.Value→ 2.8.4. Application + Infrastructure only; no public API change, so no plugin rebuild.DeepSlumberReconciler.FreeForeignSharedFactorsnow frees a cross-loadout copy only up to a real deficit (demand − target-area reset/replace refunds − free bag copies); a new Application-internalIFactorBagProbe(PandaInventoryProbe.ReadFactorBagCounts, over the existingCharSerialize.ItemPackagewalk) supplies the free-inventory counts. Elaina 2026-09-22.- Tree rebuild:
DeepSlumberService.ActivatePhaseAsyncconverges anchor activation on the prerequisite refusal5126(ErrTalentPreTalentNodeNotActivated) — each pass requeues the blocked anchors until the tree self-orders parent-before-child; a class round-trip no longer half-builds and locks the tree. Validated in-game (18 recovered 5126 across 9 rebuilds). - Reactive backstop:
DeepSlumberService.SocketPhaseAsyncfrees a foreign copy and retries when — and only when — a socket is refused7561(ErrSeasonTalentIntermediateNodeClassNumExceeded; a copy in an inactive tree counts toward the per-type limit, which a bag spare cannot clear). So a foreign copy is freed only when the bag lacks it (predictive) or the game itself refuses (reactive).DeepSlumberWriteCode.PreTalentNodeNotActivated(5126) /ItemClassNumExceeded(7561).
[2.8.3] - 2026-09-22
Section titled “[2.8.3] - 2026-09-22”2.8.3 (patch) — new character-portrait controls for mods (hide the weapon, custom portrait lighting) and a per-bar label-colour option for mod HUDs. Additive API; binary-compatible with all existing plugins (EntityInspector is rebuilt alongside).
- Mods that show a 3D character portrait (such as Entity Inspector) can now hide the character’s weapon and light the portrait with a custom, fixed look that doesn’t change with wherever your character happens to be standing. The portrait keeps its lighting entirely to itself, so characters in the world around you look exactly the same as before.
- Mods can now set the colour of the small number or label drawn inside a HUD bar, so it stays readable over any bar fill colour.
Developer notes
Section titled “Developer notes”Stellar.Abstractions:IEntityPortraitgainsShowWeapon(bool; default hidden — matches the prior always-hide behaviour) andLightPreset(int;-1= Scene/dynamic,0..4= custom profiles rebased on the real HDR scene scale). The camera controls (Orbit/Zoom/Pan/SetViewport) are factored onto a new baseIEntityPortraitView(IEntityPortrait : IEntityPortraitView) to keep each interface under the STELLAR0005 8-member cap. This is additive for consumers of the aggregate interface, but moving members to a base type changes their declaring interface — a plugin that CALLS the camera methods and is not recompiled would hitMissingMethodException; the only consumer is EntityInspector, released in tandem.- World isolation: creature lighting is a global shader register shared with in-world characters. Infrastructure (
PortraitModelHost.Lighting/PortraitCmdRenderer) captures the 5 creature-light globals (Shader.GetGlobalVector/MatrixofZModel2RTLight’s prop-id handles) before the portrait’s CommandBuffer draw and restores those exact values (cmd.SetGlobalVector/Matrix) after, so the global state ends byte-identical to what the world left it regardless of frame timing. Weapon-hide excludes the subject’s loosech_wp_*MountComp GameObjects by name from the custom draw and re-asserts on a throttle (the weapon streams in after the body). Stellar.Abstractions:BarElement.LabelColor(ColorRgba?) — a per-bar override for the inside-label colour that wins over the theme’s muted reskin registration;WindowBuilder.Previewhonours it. Defaultnullkeeps every existing bar byte-identical.FrameworkVersion.Value2.8.3. Consumed by StellarEntityInspectorPlugin (separate PR).
[2.8.2] - 2026-09-21
Section titled “[2.8.2] - 2026-09-21”2.8.2 (patch) — fixes stat readouts stuck on “—” for a whole play session after a fast login. Infrastructure/Application only; no API change, binary-compatible with all existing plugins.
- Fixed tracked stats showing “—” instead of a number for the rest of the session. If you got into the world quickly, the mod could read your character before the game had filled in its stats, decide those stats did not exist, and never look again — Stat Inspector’s mini HUD then showed a dash on every row until you restarted the game. It now waits until your stats are actually there before giving up on one, and re-ticking a stat in the picker (or logging out and back in) re-checks it straight away. A stat the game only starts reporting a little later is picked up on its own within about a minute; a stat your character genuinely does not have — one the game never reports at all — keeps showing a dash, which is correct.
Developer notes
Section titled “Developer notes”- Root cause:
PandaPlayerStatsProbe._attrUnreadablewas a process-lifetime negative cache with no invalidation. The local-player entity exists for several ticks before its FightAttr sheet is populated, so a sampling pass taken inside the login window missedTryGetAttr<long>,<int>AND<float>for every subscribed id, andProbeAndMemoizeAttrTypelatched them all.PlayerStatsService.ClearSessiondid not touch probe state andSubscribehad no reset hook, so only a client relaunch cleared it. All-or-nothing by construction — which is why the owner saw eight of eight rows dash (2026-09-21) and why the 2026-09-09 boots, which won the same race, latched only the genuinely absent 11760/11980. - Fix: the decision is now the pure, BCL-only
AttrReadabilityMemoinStellar.Application(unit-pinned inStellar.Application.Tests), one instance shared byPlayerStatsServiceandPandaPlayerStatsProbefrom the composition root. The probe reports each read as anAttrReadOutcomeinto a reused per-tick buffer andRecord(pass, sheetReady)commits the misses only when the attribute sheet is known to be populated — either because an id in the same pass read live, or because the caller’s EXPLICIT readiness signal says so. An all-miss pass before readiness means “the sheet is not ready”, never “these ids are absent”. Readiness is deliberately not inferred from the pass: the pass covers only the plugin-subscribed union, and a CombatMeter-only client subscribes exactly 11760 + 11980 — both genuinely absent from the wire sheet — so every pass there is all-miss forever and those ids would never latch, re-probing three reflective invokes each per tick for the process lifetime.PandaPlayerStateProbe.IsAttrSheetPopulated(entity)answers it from an always-present attribute (max HP, then level — the same evidenceIsUsefuluses) on the SAME entity the pass sampled, so it cannot disagree with the pass and goes false again during a mounted blackout;AttrReadabilityMemo.NeedsReadinessSignal(pass)keeps that read to the ambiguous window only (all-miss AND holding a first-read probe), so steady state pays nothing. A miss through an already-locked storage-type memo (mounted blackout) is readiness evidence only and can never latch.SubscribecallsForget(id);ClearSessioncallsClear(). OneSTELLAR_DIAGNOSTICS-gated line inPandaPlayerStatsProbe.Diagnostics.csreports each memo decision (the previous silence is what made this a full investigation). Reads stay lock-free (copy-on-write) so the per-tick sampler is unaffected.FrameworkVersion.Value2.8.2. Dossier: devkit.superpowers/sdd/statinspector-dashes-investigation.md. - An “unreadable” verdict is a BOUNDED negative cache, not a process-lifetime one:
AttrReadabilityMemo.RetryAfterTicks= 60 s._unreadableis now a copy-on-writeDictionary<int, long>(id → the clock reading that latched it);IsUnreadable(id, nowTicks)is true only inside that window, andRecord(pass, sheetReady, nowTicks)re-stamps a miss that misses again (the window restarts) or drops the verdict outright when the id finally reads — the probe’s storage-type memo (_attrPrefersLong/_attrFloat) is untouched either way.PandaPlayerStatsProbe.TrySamplereads the clock ONCE per pass (Environment.TickCount64 × TimeSpan.TicksPerMillisecond— monotonic, so a wall-clock adjustment can neither freeze nor shorten a window) and hands it down;Stellar.Applicationstays pure. Measured trigger: the 2026-09-22 TEST-client run latched11951(versatility %, a SELECTED id on that client) alongside the genuinely absent 11760/11980 — a hit pass at login condemned an attribute the game had simply not published yet, and with a permanent verdict the row dashed until relaunch. Cost of the retry: ~3 reflective invokes per absent id per MINUTE (pinned: no re-probe inside the window across 3 599 consecutive passes), versus 3 per tick without any memo. Two logging rules follow from the expiry, both required — (1) the UNGATEDattr {id} unreadable …line is now first-sighting-only per id (_unreadableLogged), else every user would collect one line per absent id per minute for the whole session; (2) theSTELLAR_DIAGNOSTICS-gatedmemo: latched=… via=…line’s expected cadence is once per id set per window, not once per process — more often means verdicts are being dropped early, and an id that stops appearing has started reading. Four new pins inStellar.Application.Tests:a_latched_id_is_re_probed_after_the_ttl_and_not_before,a_re_probe_that_still_misses_re_latches_and_restarts_the_window,the_anti_spam_bound_holds_inside_the_window,a_re_probe_that_hits_clears_the_id(PR #88 review, fix 3). - A pass taken while the attribute sheet is DARK (all-miss and
sheetReadyfalse) no longer discards its evidence wholesale:AttrReadabilityMemo.RefreshExistingrestarts the window of every fully-probed miss that ALREADY holds an entry, while an id with no entry still latches nothing (the F1 rule is untouched, andLatchedstays empty — no verdict is new). Without it the expiry introduced the mirror of the bug it fixed: an id whose window lapsed during a blackout kept its already-expired stamp, soIsUnreadableread false and the probe re-probed it — three reflective invokes plus up to three Unityarr type errlines — on EVERY tick until the sheet answered, instead of once a window. The gatedmemo: skipped-all-miss-pass …line now endslatching nothing; existing verdicts held; the refreshed ids are deliberately not listed (the line is deduped per streak, the refreshes recur per window). Three new pins:a_not_ready_pass_refreshes_an_existing_verdict_but_never_creates_one(RED before the fix),the_retry_window_ends_at_exactly_sixty_seconds(age < RetryAfterTicks, so 60 000 ms is already expired) anda_clock_reading_behind_the_stamp_reads_as_expired(PR #88 review, fix 4).
[2.8.1] - 2026-09-13
Section titled “[2.8.1] - 2026-09-13”2.8.1 (patch) — fixes a startup crash on the Steam version of the game when a mod that draws its own overlay is left enabled. Infrastructure/boot-order only; no API change, binary-compatible with all existing plugins.
- Fixed the game crashing on startup on the Steam version of Star Resonance when a mod that draws its own overlay (such as Minimal Nameplate) was left enabled. Turning it on while already in-game had always worked; only having it enabled before launch could crash. (The standalone launcher version was not affected.)
Developer notes
Section titled “Developer notes”- The 2.7.4 Steam Class::Init injection guard (
Il2CppClassInitFix.EnsureSeeded) was armed only insideUnityTickHost.Install(), whichBootstrapPlugin.OnHotUpdateReady()calls AFTERLoadUserPlugins(). That protected the framework’s own first IL2CPP type injection (StellarTicker) but not a user plugin that injects a type in its constructor. Minimal Nameplate, when enabled, callsClassInjector.RegisterTypeInIl2Cpp<LateUpdater>()in its ctor duringLoadUserPlugins— the real first injection in the process — before the seed was armed, hitting the unguardedFindClassInitand fatally crashing theStarSEA_STEAMbuild (0x80131506). Fix: callEnsureSeeded(log)immediately beforeLoadUserPlugins(log); it is_seeded-guarded/idempotent, never throws, and needs onlyGameAssembly.dllexports (available before plugins load). The existingUnityTickHost.Installcall becomes a harmless backstop.FrameworkVersion.Value2.8.1; no-op on the standalone build. Root-cause dossier: devkitdocs/recon/steam-client-classinit-injection.md.
[2.8.0] - 2026-09-09
Section titled “[2.8.0] - 2026-09-09”2.8.0 (minor) — damage-meter bar numbers can now be drawn plain, with an outline, or on a soft shadow; the meter plugin exposes the choice. Additive API; binary-compatible with all existing plugins.
- Mods can now choose how the numbers drawn on a damage-meter bar stay readable over the bar colour: plain text, the same thin dark outline the stat HUD uses, or a soft dark shadow under the numbers. CombatMeter 2.10.0 offers this as an Appearance option, useful with its new class-coloured bars (yellow, green and orange fills made plain white numbers hard to read). Nothing changes until a mod asks for it.
Developer notes
Section titled “Developer notes”Stellar.Abstractions:MeterRowData.LabelStyle(MeterLabelStylePlain = 0 / Outline / Shadow). Default Plain keeps every pre-2.8 plugin’s rows byte-identical.WindowBuilder.MeterRowbuilds all three treatments per row and the binding toggles them from the row data: Outline = the framework’s existing readability outline (TextElement(Shadow: true)’s recipe, StatInspector’s stat HUD: uGUIOutline, black @ 0.85,effectDistance (1.1, -1.1); factored intoUGuiPrimitives.AddReadabilityOutlineand shared by both call sites — the 1.1 px sub-pixel offset is what keeps a small bold glyph’s halo smooth; exactly 1.0 px stair-steps); Shadow = twoRawImagefades under the values (64 px × font scale, plateau + SmoothStep ramp, mirrored for the right end) whose alpha lerps 0.12 → 0.52 with the fill’s relative luminance (LabelFadeTint, retinted whenever the fill colour changes). Measured trigger: with CombatMeter 2.10.0’s class palette the fill can be#f2dc3a/#5ccf36/#f5761c, where white text has a WCAG contrast of 1.4–2.0. Owner 2026-09-09 after four in-game rounds: “it should be option for players: normal text, outline, smooth shadow”.FrameworkVersion.Value2.8.0. Design record: devkitdocs/superpowers/specs/2026-09-09-game-class-palette-and-gauge-color-design.md.
[2.7.4] - 2026-09-09
Section titled “[2.7.4] - 2026-09-09”2.7.4 (patch) — fixes a mod-caused startup crash on the Steam version of the game. Infrastructure-only; no API change, binary-compatible with all existing plugins.
- Fixed the mod crashing the game on startup on the Steam version of Star Resonance. (The standalone launcher version was not affected.)
Developer notes
Section titled “Developer notes”- Il2CppInterop’s IL2CPP type injector (
InjectorHelpers.FindClassInit) finds nativeClass::Initby scanningGameAssembly.dllfor a hardcoded x86-64 call-site byte-signature. The SteamStarSEA_STEAMbuild — a separate SKU compile, identical Unity 2022.3.59f1 / IL2CPP metadata v31.1 — has codegen that scan mis-resolves to a bad pointer, hard-crashing the CLR (0x80131506, an uncatchableExecutionEngineException) at the framework’s FIRSTClassInjector.RegisterTypeInIl2Cpp(StellarTicker). Reproduced identically on Il2CppInterop 1.5.1 and 1.5.3, so a BepInEx bump does not fix it. - Fix: new
Il2CppClassInitFix.EnsureSeeded(Infrastructure), called at the top ofUnityTickHost.Installbefore the first injection, reflectively pre-seeds the process-globalInjectorHelpers.ClassInitwith Il2CppInterop’s own documented substitute export (il2cpp_class_has_references) so theClassInit ??= FindClassInit()guard skips the broken scan. Always-on (all builds): harmless where the scan already worked — the substitute is the library’s own fallback for injected classes — verified in-world on the standalone with all injections (StellarTicker,WindowInteractionTicker,ChartGraphic) succeeding. Idempotent, never throws.FrameworkVersion.Value2.7.4; no API change. Root-cause dossier: devkitdocs/recon/steam-client-classinit-injection.md.
[2.7.3] - 2026-09-09
Section titled “[2.7.3] - 2026-09-09”2.7.3 (patch) — stable release of the 2.7 line. Identical code to the 2.7.2 testing build; the number moves only because 2.7.2 was already published to the testing channel and a taken version is never re-minted. No API change; binary-compatible with plugins built against ≤2.7.2.
- Stable release of the 2.7 line. Same code as the 2.7.2 testing build — the buff and stat recording the logs site’s experimental raid-contribution view needs (Own / Given, and the Buffs tab), plus everything from 2.6.4. Nothing else changes for players coming from 2.6.4 beyond that recording.
Developer notes
Section titled “Developer notes”- Promotion only — no source change. The merged tree is byte-identical to
feat/attr-changed-event@73f349e(the build on the testing channel since 2026-09-09);origin/main(2.6.4,d6a51db) contributed no content because the branch had already merged thev2.6.4tag lineage.FrameworkVersion.Value2.7.3. - 2.7.2 cannot be re-used for the stable publish:
Stellar-2.7.2.zipis already on the CDN under the bundle-immutability guard, and a GitHub pre-release taggedv2.7.2already exists — so the stable promotion rides a patch bump. What the 2.7 line actually adds is under the 2.7.2 / 2.7.1 / 2.7.0 entries below.
[2.7.2] - 2026-09-09
Section titled “[2.7.2] - 2026-09-09”2.7.2 (patch) — the testing-channel framework now carries the 2.6.3 and 2.6.4 fixes. Additive; binary-compatible with plugins built against ≤2.7.1.
- Mods can now tell which of your characters you are currently playing, so features that save things per character keep each character’s data separate.
- Fixed moved HUD elements (quest tracker, boss health bar, and others) jumping to the wrong position after you change your screen resolution.
- Fixed the boss health bar drifting out of place after you repeatedly hide and show it in Settings.
- HUD elements you have repositioned now fall back to their normal in-game position at a resolution you have not customized yet, instead of getting stuck off to one side.
Developer notes
Section titled “Developer notes”- Merge of
v2.6.4(d554236, built onv2.6.3) into the 2.7.x line (feat/attr-changed-event, on top of 2.7.1 = 2.7.0 + v2.6.2);FrameworkVersion.Value2.7.2. Bringslong IPlayerIdentity.CharIdand the native game-UI repositioning fixes; the 2.7.0CombatEvent.EntityAttributesChangedgroundwork is unchanged. Details under the 2.6.4 / 2.6.3 entries below.
[2.7.1] - 2026-09-08
Section titled “[2.7.1] - 2026-09-08”2.7.1 (patch) — the testing-channel framework now carries the 2.6.2 loadout fix. No new API; binary-compatible with plugins built against ≤2.7.0.
- Switching between two loadouts that share a Deep-Slumber (Psychoscope) factor no longer stops partway with “Deep-Slumber partly applied” — the 2.6.2 fix, now included in the testing channel.
Developer notes
Section titled “Developer notes”- Merge of the
v2.6.2hotfix (hotfix/deepslumber-cross-loadout-2.6.2, off v2.6.1) into the 2.7.0 line (feat/attr-changed-event);FrameworkVersion.Value2.7.1; no source change beyond the merge and the version constant.
[2.7.0] - 2026-09-05
Section titled “[2.7.0] - 2026-09-05”2.7.0 (minor) — no player-visible change yet; internal groundwork for an upcoming raid-contribution DPS (rDPS) feature. Additive, binary-compatible with plugins built against ≤2.6.1.
Developer notes
Section titled “Developer notes”- CombatEvent.EntityAttributesChanged: a player’s numeric attributes changed (one event per wire packet, stamped like the packet’s buff events) — enables the CombatMeter stat-sheet track for rDPS.
- Entity attribute capture now stores a genuine zero value (single 0x00 varint) instead of dropping it; non-varint payloads are still skipped.
[2.6.4] - 2026-09-09
Section titled “[2.6.4] - 2026-09-09”2.6.4 (patch) — new plugin API so mods can tell which of your characters is logged in. Additive interface member; binary-compatible with plugins built against ≤2.6.3. Requires updating LoadoutSwitcher + Wardrobe to their per-character builds to take effect.
- Mods can now tell which of your characters you are currently playing, so features that save things per character keep each character’s data separate. This makes the LoadoutSwitcher and Wardrobe updates possible: your Deep-Slumber loadout bindings and your saved outfits are now remembered per character instead of one shared set that switching character would overwrite.
Developer notes
Section titled “Developer notes”- Adds
long IPlayerIdentity.CharId(Stellar.Abstractions/Services/IPlayerState.cs): the local character’s stable id, char-record-backed so it is known before the world entity syncs (likeName/Level/Profession),0when not yet resolved; NOT gated byIsAvailable.PlayerStateServiceexposes the already-tracked_identityCharId(fed fromIPlayerStateProbe.CharId, cleared to 0 on logout / character change) — no probe change.MockPlayerStatereturns a fixed id for tests/sandbox. - Additive (a new interface member + getter): binary-compatible with plugins built against ≤2.6.3, but a plugin that calls
CharIdrequires this framework — so LoadoutSwitcher and Wardrobe bump theirminframework to 2.6.4. Abstractions/Application surface only; built on v2.6.3 (native-UI layout fixes), which also carries the 2.6.2 Deep-Slumber cross-loadout factor-move fix. - Pinned by
PlayerStateServiceidentity testCharIdReflectsIngestedIdentityAndResetsOnLogout(CharIdreflects the ingested identity, resets to 0 on logout / character change).
[2.6.3] - 2026-09-08
Section titled “[2.6.3] - 2026-09-08”2.6.3 (patch) — moved HUD elements stay where you put them across resolution changes and Settings toggles. Infrastructure/Application-only, binary-compatible with plugins built against ≤2.6.2.
- Fixed moved HUD elements (quest tracker, boss health bar, and others) jumping to the wrong position after you change your screen resolution.
- Fixed the boss health bar drifting out of place after you repeatedly hide and show it in Settings.
- HUD elements you have repositioned now fall back to their normal in-game position at a resolution you have not customized yet, instead of getting stuck off to one side.
Developer notes
Section titled “Developer notes”- All in the native game-UI repositioning path (
PandaHudAdapter/NativeUiService/INativeUiAdapter). TheSetRectidempotent guard was reworked to key on(target, anchoredPosition, curatedSize)equality, fixing three defects plus a new position-only fallback: (1) reflow-freeze — a moved HUD element dropped to the bottom after a resolution round-trip (translate applied against a transient/collapsed size, then froze when the element reflowed); (2) toggle-ratchet — the boss HP bar crept left on repeated hide/show in Settings; (3) off-edge churn — elements whose curated rect sits past a screen edge re-applied every tick (harmless, now skips); (4) unconfigured-resolution fallback — newINativeUiAdapter.RestoreOriginalPosereturns a moved element to the game’s default position at a resolution with no saved layout (previously it stayed stranded at the prior resolution’s offset). Scope:src/Stellar.Infrastructure/Game/PandaHudAdapter.cs,src/Stellar.Application/Services/NativeUiService.cs,src/Stellar.Application/Abstractions/INativeUiAdapter.cs,tests/Stellar.Application.Tests/NativeUi/NativeUiServiceTests.cs.INativeUiAdapteris an Application-layer outbound interface, not the public plugin surface — no plugin API or behavior change outside native-UI positioning. Compile clean; 1394/1394 tests pass (incl. a new unconfigured-resolution fallback test); all four cases in-game validated by the owner. Temporary round-trip diagnostics were added during development and stripped before release (repo grep clean).
[2.6.2] - 2026-09-08
Section titled “[2.6.2] - 2026-09-08”2.6.2 (patch) — switching loadouts now moves a shared Deep-Slumber factor instead of half-applying. Application-only, binary-compatible with plugins built against ≤2.6.1.
- Switching between two loadouts that share a Deep-Slumber (Psychoscope) factor no longer stops partway with “Deep-Slumber partly applied”. Because a factor exists only once, if the loadout you are leaving is still holding one the new loadout needs, the switch now takes it out of the old loadout first and puts it into the new one — so the switch completes. Nothing is reset and it costs no season points; the factor simply moves, the same as if you had moved it by hand.
Developer notes
Section titled “Developer notes”DeepSlumberReconcilergains a foreign-free pass (FreeForeignSharedFactors): after the per-target-area reconcile, any factor the target wants that is still socketed in a CURRENT-season area the target does not bind is unsocketed (into the existing enable→reset→unsocket→activate→socket phase, ahead of every socket), so the target socket finds it in the bag instead of failing game code 7561 (“still socketed elsewhere”). Scoped to the current season (currentLine), to areas OUTSIDE the target (AreaIdnot bound — the per-area logic already reconciles the target’s own areas, so re-touching them would double-emit), and to items the target actually wants (never disturbs the other loadout’s non-shared factors); deduped by (AreaId, node) so a node id repeated across different foreign areas never drops a wanted factor. Unsocket is free + non-consuming, so this never spends reset currency. Confirmed in-game on the owner’s MAIN (2026-09-07):worldProxy.UnInstallItemToMiddleNode({nodeId})returns code 0 against a NON-active line’s node — unsocket needs no active line, superseding the 2026-08-24 “area-relative” inference — and a real Smite↔Tank switch (12 factors) reads[LoadoutSwitcher] DS apply -> Success, with every DS op code 0 and no 7561/7555. Pinned byDeepSlumberReconcilerTests(cross-loadout free, two-foreign-areas-same-nodeid, replace+foreign coexistence, prior-season-not-freed, not-wanted-left-alone). Toir report via the owner, 2026-09-07.
[2.6.1] - 2026-09-05
Section titled “[2.6.1] - 2026-09-05”2.6.1 (patch) — “no weapon skin” is saved as no weapon skin again. Infrastructure-only, binary-compatible with plugins built against ≤2.6.0.
- Saving an outfit while your weapon wears no skin now records “no skin”. It used to record whichever skin the wardrobe was showing you at the time, so applying that outfit later put that skin back — even after you had changed to a different weapon, and even though you had never chosen a skin. Now an outfit saved with no weapon skin always shows your current weapon’s own look, whatever weapon you are holding. Outfits you already saved keep the skin they recorded until you use Update on them.
Developer notes
Section titled “Developer notes”- Capture-side half of the
SkinId 0⇄ origin contract. The game has no “skin 0”: the Weapon Skin tab builds its tiles withisEmpty = value.Original == 1and marks worn the tile whoseId == weaponSkillSkinVm:GetWeaponOriginSkinId(curProfessionId)(fashion_weapon_skin_select_view.lua:116-120), so picking ⊘ sendsUseProfessionSkin(pid, <that origin row id>)and the server stores a CONCRETEWeaponSkinTablerow whose look happens to be the weapon’s native one. Measured on the owner’s report: stored outfit 9 heldWeaponProfessionId=5, WeaponSkinId=7350002= “Mirrorlight Ring”,ProfessionId 5,Original: 1— and profession 5 has fourOriginal==1rows (7350002, 7350006, 7351061, 7351062), one per weapon family, so the raw id pins the outfit to the weapon it was saved with.PandaFashionProbe.WeaponCaptureLuanow normalises a worn skin equal toGetWeaponOriginSkinId(cur)back to0, wherecurisCharSerialize.professionList.curProfessionId— exactly the idprofession_vm.GetContainerProfessionreturns, which that VM method requires to match or it returns nil. Both consumers already resolved0→ origin at USE time (BuildWeaponSkinChunkfor apply, mirroringAsyncUseProfessionSkin:199;PandaWardrobePreviewProbe.WeaponSkinSourceLuafor the 3D preview), so only the producer was wrong;ParseWeaponSkinand the"<pid>:<skin>"global format are unchanged, and theIWardrobe.GetWornWeaponSkinXML doc now states why0is the honest value. The origin lookup sits in its OWNpcallnested inside the weapon block’s, so a nil VM, a throwingGetVM, a throwingGetWeaponOriginSkinIdor an origin that resolves0all keep the RAW id rather than losing the capture. Verified against real Lua 5.3 with stubbedZ.ContainerMgr.CharSerialize+Z.VMMgr.GetVM('weapon_skill_skin'), 9/9 cases: worn 7350002 / origin 7350002 →"5:0"; worn 7350003 / origin 7350002 →"5:7350003"; origin call throws →"5:7350002";UseSkinIdnil →"5:0"(no VM call at all); plus VM-nil, GetVM-throws, origin-0, float-typedUseSkinIdand no-current-class. Pinned byWardrobeWeaponSkinTests.CaptureChunk_ReportsTheCurrentWeaponsOriginSkin_AsNoSkin. Owner report 2026-09-05, after the 2.6.0 release.
[2.6.0] - 2026-09-05
Section titled “[2.6.0] - 2026-09-05”2.6.0 (minor) — saved outfits can now carry your weapon skin, and the outfit preview shows both head accessories. Additive, binary-compatible with plugins built against ≤2.5.0.
- Plugins can now read the weapon skin your current class is wearing and switch it for you, through the same game action as the Wardrobe’s Weapon Skin tab. The Wardrobe plugin uses this to save and re-apply your weapon skin together with an outfit.
- The outfit 3D preview now shows the weapon skin a saved outfit carries, so you see the whole look before you switch. Weapon skins belong to one class, so the preview shows it when the outfit’s weapon skin is for the class you are currently playing — otherwise the weapon is left as it looks now.
- The outfit 3D preview no longer shows two weapons at once. When a saved outfit carried a weapon skin, the preview drew that skin and the weapon you are currently holding; it now shows only the saved one, the way the game’s own wardrobe and shop previews do. Outfits saved before weapon skins existed still preview with the weapon you are wearing.
- The outfit 3D preview now shows both head accessories at once. An outfit with two head pieces used to preview only the second one — the first appeared for a moment and was then replaced.
- No more freeze when a plugin saves its settings. Saving an outfit in the Wardrobe used to lock the game up for about a third of a second and then stutter for a few seconds afterwards — the more you had saved, the worse it got. Saving is now instant, however many outfits you keep.
- Switching to the loadout you are already wearing no longer re-equips everything. Pressing it a couple of times in a row used to freeze the game for seconds and then stutter; it now does nothing but tell you that loadout is already on.
- On-screen notices no longer hitch the game when several arrive at once. Saving a few outfits in a row, or tapping a loadout hotkey repeatedly, used to drop frames for every notice it put on screen; notices now reuse the banner that is already up, and a notice repeating word-for-word while its own copy is still visible is left alone instead of stacking.
- Loadout switches now respect the game’s own three-second switch cooldown, exactly like the in-game dropdown does: a second press right after the first is refused — with the game’s own “you switched too recently” message — instead of piling another full re-equip on top of the one still running.
Developer notes
Section titled “Developer notes”IWardrobegainsGetWornWeaponSkin()→WardrobeWeaponSkin(ProfessionId, SkinId)(read fromCharSerialize.professionList.professionList[curProfessionId].UseSkinIdin the same capture chunk as the outfit) andApplyWeaponSkinAsync(professionId, skinId)(drivesWorldProxy.UseProfessionSkinexactly asweapon_skill_skin_vm.AsyncUseProfessionSkindoes: skin 0 → the class’s origin skin,OnWeaponSkinChangedispatch on ok; shares the single in-flight slot withApplyAsync, so await the outfit before sending the skin). Weapon skins stay OUT of the outfit region map (WardrobeRegions.Allunchanged) — they are a per-class game system. Pinned byWardrobeWeaponSkinTests+WardrobeServiceTests.PandaWardrobePreviewProbenow stamps eachSingleWearData.SlotIDwith the piece’sFashionRegion— the shape of the game’s ownfashion_vm.GetFashionWearList(data.SlotId = region). SlotID routes head pieces to their mount (713 → HeadWear, 718 → HeadWear2);SlotID=0put both on one mount so the second overwrote the first. Pinned byWardrobePreviewChunkTests. Discord report 2026-09-03.- Weapon-skin preview:
WardrobeRegions.WeaponSkinPreview(731) is a preview-only key onIWardrobePreview.Show’s outfit map —IWardrobe.ApplyAsyncignores it and it stays out ofWardrobeRegions.All(still 14). A weapon skin can never travel inEWearFashion(the game’s ownfashion_vm.GetFashionWearListskips 731), soPandaWardrobePreviewProbekeeps 731 out of theSingleWearDatalist and instead re-skins the social data the model is built from:socialData.professionData.weaponSkin = skin, written beforeZ.ModelManager:GenModelByLuaSocialData.SocialData.professionDatais the{profession_id, weapon_skin}block the server fills forSocialDataType.SocialDataTypeWeapon(const_value.lua:117), andsocialDatais a plain Lua table (world_proxy.GetSocialData:21831returnspb.decodeoutput with camelCase keys), so the field is directly writable. The display-override attr (m:SetLuaIntAttr((Z.ModelAttr).EModelDisplayWeaponSkinId, skin)) is still emitted last as a secondary, but it is NOT the mechanism: all three game call sites dress the cached PLAYER model (GetCachePlayerModel—fashion_system_view:831,shop_fashion_sub_view:300,competency_rating_main_view:253), never a social-data model, and the owner’s in-game test showed the preview keeping the live weapon for every outfit. The injection alone renders TWO weapons (owner, 2026-09-05: “it show with currently weapon using … previewer show player have 2 weapons skin rendered”, withsocial=set(prof=2,was=nil)in the log): a social-data model carries two independent weapon renderers — the CMount attachment (EModelCMountWeaponL/R, string model paths that ride in with the social data;Panda.ZGame.WeaponOriginDatabundlesModelCMountWeaponL/RbesideWeaponSkinId/MainModelId) andPanda.ZGame.WeaponModelComp, which owns its own weapon GameObjects (mainWeaponModel_/subWeaponModel_,getWeaponMount/clearModel) and is driven by the profession+skin data. So the chunk now does what every game view that shows a weapon skin on a preview model does — clear the CMount first, then ask for the skin:fashion_system_view.initPlayerModel:844-845(the screen hosting the weapon-skin tab whoseSelectStyle:26sets the override) andshop_fashion_sub_view.initPlayerModel:319-320(ShowPlayerWeaponModel:275sets it), the sameSetLuaAttr(EModelCMountWeaponL/R, "")pair every weaponless social view uses (investigation_clue_window_view:942/944,rank_main_view:503/637). The clear is emitted only when the outfit carries a weapon skin and is skipped at run time when the resolved skin is0, so no outfit can leave the model unarmed. Skin0= the class’s default look and resolves throughGetWeaponOriginSkinIdexactly asAsyncUseProfessionSkin:199does. Class-match is the caller’s rule (everyWeaponSkinTablerow isProfessionId-scoped; the game’s own tab refuses other classes atfashion_weapon_skin_select_view.lua:100-106). All three weaponpcalls CAPTURE their result and every preview writes a one-line outcome the host logs unconditionally:[WardrobePreview] weapon none/… weapon skin=<id> social=set(prof=2,was=nil) mount=cleared(L:<path>,R:<path>,skinModels:<WeaponSkinTable.WeaponModelId list>) disp=ok/… mount=skip(skin0)/… social=err:|mount=err:|disp=err:. The mount values are read BEFORE the clear, in their own nestedpcallso a read failure can never block the clear, and the skin’s own weapon model ids sit beside them — one owner log now says which renderer held which weapon. The 2.6.0 silentpcallmade the failure undiagnosable from the owner’s log. Pinned byWardrobePreviewChunkTests. Owner reports 2026-09-05. - Plugin-config save path taken off the large-object heap. Owner report 2026-09-05: every click of Wardrobe’s save button (53 outfits, 57 KB config) froze a frame for 287-366 ms and then stuttered for seconds. Measured on the owner’s real config, 53 saves × 3 watcher events, running the actual
PluginConfigService+FileConfigStore: 1,169 KB/save and 7 full gen2 GCs → 130 KB/save and 0 GCs of any generation (CPU 1.67 → 0.64 ms/save; gen2 pauses on this client are 100-475 ms, per the 2026-07-25 jitter investigation). Three causes, all fixed: (1)FileConfigStore.HandleFileTouchdidFile.ReadAllText+Encoding.UTF8.GetBytes+ SHA256 for EVERY FileSystemWatcher event (Wine delivers 2-4 per write) before the self-write check — a 106 KB LOH string + 54 KB array just to recognise our own write; a newSelfWriteLedgernow answers that from oneFileInfo((Length, LastWriteTimeUtc)) with the content hash kept as the fallback for events that raceFile.WriteAllText. (2)PluginConfigService.SaveSectiondeep-cloned the whole root (ToJsonString+Parse) only to hand it to a store that serializes it again — it now passes the live root under the existing lock, andIConfigStore.Savedocuments the serialize-synchronously / retain-nothing contract. (3) Config files are now written compact, not indented — the owner’s nested config inflated 20 K → 55 K chars when indented, putting every save’s string over the 85 KB LOH threshold; readability of the on-disk file is the deliberate trade, and reading still accepts indented files (rollback-safe, process rules § 6). Pinned bySelfWriteLedgerTests(11),FileConfigStoreEchoTests(9, incl. zero-read echo suppression and a same-length external edit that must still be detected) andPluginConfigServiceTests.Save_HandsTheStoreTheLiveRoot_NotAFreshClonePerSave. WindowBuilder.LoadIconnow dedups Texture2D decodes through the token’s byte[]-keyedAtlasCache— previously onlySpriteElementand the live tile-icon binding did, so every other icon leaf (button chips, images, brand logo, pin stars) re-decoded per widget: the Wardrobe’s 20-row × 9-chip pool built ~180 textures from 8 distinct PNGs on each window open. Dedup now lives in exactly one place (BuildSpriteandIconBindingdropped their own copies). Safe to share because tinting is per-graphic (.color) and sub-rect selection per-RawImage(.uvRect) — nothing is baked into the texture;IconTexturesremains the single owner, so disposal is still oneDestroyper texture.- Loadout switch pre-dispatch gates (owner report 2026-09-05: three same-loadout presses = a 3,261 ms frametime spike, then stutter; log showed eight consecutive
[LoadoutSwitcher] Switched to Beam). The removed fast-path comment inPandaLoadoutProbe.CallApplyAsyncclaimed the game “cheaply no-ops a switch to the already-active loadout” — REFUTED byweapon_vm.lua:509-514, which builds{oldProjectId=CurPlanId, newProjectId=planId}and firesWorldProxy.SwitchProjectunconditionally, then (:536-542) stampsSwitchRolePlanTime, savescurrentProjectSyncDataand dispatchesOnRolePlanChange— a full server re-equip. The game’s own dropdown never allows it:role_plan_loop_item.lua:66ignores the click whenCurPlanId == PlanId,:118-121refuses insideGlobal.lua:1703 CombatStrategySswitchCd = 3(tip 150208),:122-125refuses in battle (tip 150206). New purePandaLoadoutProbe.DecideSwitchmirrors the first two: same plan →LoadoutResult.Successwith nothing dispatched ([Stellar][Loadout] switch to <id> skipped: already active), within 3 s of our last dispatch →Rejected(… refused: within the game's 3s switch cooldown). Combat is unchanged — the server + the game’s own wrapper still own that refusal.PendingSwitchnow records whether the live plan already matched at dispatch and skips the vacuous_currentId == TargetIdcompletion fallback in that case, so a switch can no longer “complete” on the first ~33 ms poll and let the caller’s single-flight guard start an overlapping burst. Pinned byPandaLoadoutProbeSwitchGateTests. A cooldown refusal also shows the game’s OWN tip rather than only a log line (the LoadoutSwitcher plugin’sReporttoasts only onSuccess):BuildShowTipsChunkemitspcall(function() Z.TipsVM.ShowTips(150208) end)— the exact dot callrole_plan_loop_item.lua:119makes, localized by the game — queued on the probe’s existing main-thread dispatch queue and drained by the sameDrainPendingDispatchespass (no new thread hop). One tip per refused press, capped atMaxQueuedTipsand dropped onClearSession. - The per-class gear/module resolve is now debounced on the TRAILING edge of the delta burst. The walk is whole-item-container (uuid index over every package + a gear read per slot per plan) and a re-equip burst re-arms it on every
CharSerializedelta, so ungated it ran at the full ~30 Hz drain rate through each burst — the ~1.8 MB/s allocation class behind the 2026-07-25 A/B’s 100-475 ms GC frames. A leading-edge cooldown (walk now, then wait a window) bounded the rate but still walked THROUGH the burst, twice for a ~1 s burst, both times against half-applied state.TryResolvePerClassDetailsIfDuenow consumes_resolvePendingonce per tick into a sticky_resolveArmed(the arming sites live in sibling partials and a bool cannot carry “again”), restarts a quiet timer on every new arm, and walks only afterResolveQuietTicks(15 ≈ 0.5 s) of silence — one loadout switch = ONE walk, against the burst’s final state. DEFER, NEVER DROP: the arm survives in_resolveArmeduntil the walk runs, andResolveMaxDeferTicks(60 ≈ 2 s, counted on EVERY armed tick so a saturating stream cannot starve it) forces a walk during an endless stream, so “late, never stale” keeps a hard ceiling. The purePandaLoadoutProbe.DecideResolvenow takes(resolveArmed, resolverAttached, hasInputs, quietTicks, deferTicks); the gate still wrapsTryResolvePerClassDetailsrather than living inside it, so every existing live-state / Deep-Slumber resolve pin drives the unchanged inner method. New always-on, self-limiting line[PerClassLoadout] walk {ms}ms plans={n}, emitted only above 33 ms (zero volume in normal play, no diagnostics restart needed to attribute a reported spike). Pinned byPandaLoadoutProbeResolveGateTests(10, incl. a burst replay that scores 1 walk where the leading-edge form scored 2, and a saturating-stream starvation pin). - Notice tips no longer re-open
noticetip_popper tip. Owner reports 2026-09-05: (1) holding a loadout hotkey on the already-active loadout gave 4-5 frames of 100-185 ms over 3-5 s even though the framework SKIPPED every switch (no RPC) — the toast was the only game-engine touch left on that path; (2) repeated “Save current outfit” still stuttered (max 428 ms), and those toasts carry DISTINCT text so a dedupe alone could never have fixed it. Root cause:Z.UIMgr:OpenView('noticetip_pop')once per tip, run synchronously on the main thread inside the service tick. On an ALREADY-OPEN view that is not cheap (ui_manager.lua:112-161): the open list is re-ordered (:137-145),Z.UICameraHelper:OpenUICameraruns (:151), andui:Active(:152→ui_base.lua:45-59) reachesSetAsLastSibling(:53→ui_view_base.lua:80-85= a transform re-parent PLUSZ.UIMgr:UpdateDepthover the layer, i.e. a canvas rebuild) before it ever gets toOnRefresh(:55), thenViewStatusSwitchMgr:TrySetStateActive(:158) and a globalEventMgr:Dispatch(UIOpen)(:160) fire. The view already drains its own queue —OnRefreshdequeues one item when it has noviewData(noticetip_pop_view.lua:80-86),showPopTipholds three at a time and re-enqueues the overflow (:130-133), each item’sOnEndpulls the next (:210-214) — soBuildPopChunk/BuildPopTipChunknow enqueue and then refresh the LIVE view (GetView+SetViewData(nil)+CallLifeCycleFunc(OnRefresh), the same tailActivewould have reached), keepingOpenViewonly as theelsefallback for when there is no active/loaded/visible view — so no tip can be stranded in the queue. Two service-side bounds on top: the drain takes ONE tip per tick (a lone tip still shows on the very next tick), and a chunk byte-identical to one still inside its own delay+duration window is dropped by the pureNoticeTipService.DecideShow— the same call the game’s owncheckConfigRepeatmakes (noticetip_data.lua:8-22), which never fires for us because our tips carryId=0and so have no MessageTable row. DIFFERENT content is never dropped. New always-on, self-limiting line[NoticeTips] slow show {ms}ms, emitted only above 33 ms. Pinned byNoticeTipSpamTests(10).
[2.5.0] - 2026-09-01
Section titled “[2.5.0] - 2026-09-01”2.5.0 (minor) — the Loadout Switcher can switch your whole Deep-Slumber (tree and all), plugins get accurate raid boss health, and raid runs record as one run. Additive, binary-compatible with plugins built against ≤2.4.1.
- Plugins can now read a boss’s exact health straight from the game, and can tell when a raid boss has actually gone down even when the game never shows its health hitting zero. The Combat Meter uses this for accurate boss health in raid replays.
Changed
Section titled “Changed”- The Loadout Switcher now switches your entire Deep-Slumber when you change loadouts — its tree, not just the phantom factors. Switching to a loadout whose Deep-Slumber uses a different tree now rebuilds the tree for you instead of stopping half-done, and it no longer stutters while it applies.
- Raid runs are recorded as one continuous run again, instead of being split into pieces when the game’s own timers reset partway through the fight.
Developer notes
Section titled “Developer notes”DeepSlumberAreaBindinggains a non-positionalNormalNodesinit member (the tree / Anchor allocation — kept off the primary constructor so it stays binary-compatible with ≤2.4.0).DeepSlumberReconcilernow diffs the tree and, when it differs, drivesResetAllNodes+ActiveNormalNode(whole-area reset — the game has no per-node anchor removal, owner ruling 2026-09-01) before re-socketing factors; legacy bindings with a null tree stay factor-only. Both new worldProxy RPCs were validated in-game (code=0,zoneId=areaId). New Kind-phase order: enable → reset → unsocket → activate → socket.- The loadout live-state refresh now defers its full-container walk while a plugin Deep-Slumber apply is in flight (
PandaLoadoutProbe.DecideRefreshgated onPandaSeasonTalentProbe.HasPendingWrites), collapsing the per-opCharSerializeburst into one refresh after the apply settles — removes the 3-5 frame hitches a tree rebuild otherwise caused. Mirrors the existing combat-defer; a manual in-game switch is unaffected. - Native boss-HP tap:
IBossVitals+EntityVitalsServiceread boss HP from the game’s own entity, withEDisappearType-aware AOI eviction and anAttrMaxHpTotal(11321)fallback; a raid boss whose HP never reads 0 is detected as HP≈1%-then-vanish.IRunTimerlatch-epoch gives upgrade-proof run identity so an evidence-less belt resolution cuts a segment without resetting the run’s identity (raid run-split fix; spec 2026-08-26).
[2.4.1] - 2026-08-29
Section titled “[2.4.1] - 2026-08-29”2.4.1 (patch) — stops a repeated combat popup during long fights. No API change; binary-compatible with all existing plugins.
- No more repeated “Cannot perform this action during combat” popup during long fights. It could appear every few seconds in sustained combat, even with no plugins installed.
Developer notes
Section titled “Developer notes”PandaLoadoutProbenow defers only the combat-gatedSyncProjectListRPC (AsyncGetRolePlanData) while the local player is in combat, using the game’s own in-combat check (GetLuaLocalAttrInBattleShow()/GetLuaIsInCombat(), fail-safe to not-in-combat), and fires exactly one refresh at combat end; the RPC-free live-state re-read keeps running in combat. Root cause: every CharSerialize merge re-armed_refreshPending, and the server rejects the RPC in combat (ErrStateIllegal 3202), which the game’s own wrapper toasts (~every 5s as deltas drip in). Infrastructure-only (PandaLoadoutProbe,PandaLoadoutProbe.Resolution); no API change — binary-compatible with all existing plugins. +7PandaLoadoutProbeRefreshGateTests. (#72)
[2.4.0] - 2026-08-25
Section titled “[2.4.0] - 2026-08-25”2.4.0 (minor) — plugins can now save your outfits and switch between them, with a live 3D preview. Additive, binary-compatible with plugins built against ≤2.3.0.
- Plugins can now save the outfit you’re wearing and switch you back to it later, and show a live 3D preview of a saved outfit on your own character. The new Wardrobe plugin uses this for instant hotkey outfit switching.
Developer notes
Section titled “Developer notes”- New plugin surface:
IWardrobe(capture the worn outfit as a region→fashionId map; apply through the game’s ownWorldProxy.FashionWear, keeping every server-side check) andIWardrobePreview(dress a fresh self-model with an arbitrary saved outfit viaGenModelByLuaSocialData+SetLuaAttr(EWearFashion, …), rendered through a secondPortraitModelHost; orbit / zoom / pan).FashionEntrygainsDyeAreas— parallel toDyes, carrying each dye’sEFashionColorAreaTypeso multi-area pieces preview on their real areas;AttrFashionDataReadernow reads both the base (field 2) and attachment/socks (field 3) colour maps. RE:docs/recon/wardrobe-fashion-preview.md.
[2.3.0] - 2026-08-25
Section titled “[2.3.0] - 2026-08-25”2.3.0 (minor) — plugins can now apply a Deep-Slumber setup for you, and applying one is fast and self-healing. Additive, binary-compatible with plugins built against ≤2.2.0.
- Plugins can now change your Deep-Slumber Psychoscope for you — its cultivate line and its phantom factors. The Loadout Switcher uses this to re-apply the Deep-Slumber you bound to a loadout the moment you switch to it.
Changed
Section titled “Changed”- Applying a Deep-Slumber setup is much faster. A switch that moves a lot of factors used to take a few seconds; now it finishes in a fraction of that, and if the game drops one of the changes it quietly retries instead of stopping half-done.
Developer notes
Section titled “Developer notes”- New plugin surface:
IDeepSlumber.ApplySetupAsync(DeepSlumberSetup, CancellationToken)returningDeepSlumberApplyResult, plusDeepSlumberSetup/DeepSlumberAreaBinding. The live→target diff (DeepSlumberReconciler) lives in Application; the write path drives the game’sseason_talentworldProxy RPCs (Approach A — raw RPC returns the bareEErrorCodeinline) viaPandaSeasonTalentProbe. Docs:docs/driving-game-actions.md§ Deep-Slumber. - Apply overlaps its server round-trips: a bounded in-flight window (5) with one-dispatch-per-tick pacing, run in Kind-phases (enable → unsocket → socket) with a barrier that preserves the scarce single-copy unsocket-before-socket invariant. Only transient (did-not-land) ops retry (initial + 2, 250 ms backoff); a positive game refusal (7555/7561/combat) is never retried.
LauncherEntry.TitleProvider(Func<string>?): an optional live-localized tile title. The launcher rendersDisplayTitle(TitleProvider?.Invoke() ?? Title) each frame, so a tile whose plugin setsTitleProvider = () => Localization.T(key)re-localizes immediately on a language change;Titleremains the stable pinned-state identity. Plugins pass a captured string today, which is why their tiles were stuck in the registration-time language.
[2.2.0] - 2026-08-24
Section titled “[2.2.0] - 2026-08-24”2.2.0 (minor) — the live-build release. Stellar now notices the moment your setup changes and reads it live, and exposes the Deep-Slumber Psychoscope to plugins. Additive, binary-compatible with plugins built against ≤2.1.0.
- Stellar now notices the moment you change your build — swapping a piece of gear (including with the Replace button), changing modules, respeccing talents, switching a Battle Imagine, or editing your Psychoscope. Plugins that record your setup, such as the Combat Meter, can now save the exact build you fought each boss with instead of whatever you happened to have on when the run started.
- Your Psychoscope (Deep Slumber) is now available to plugins — season level, lines, socketed cards and node levels, read live from the game rather than from a saved profile.
- Plugins can now show a second value on the same on-screen bar — for example a boss’s shield alongside its health, either as a see-through band laid over the health fill or as extra length past the end of it. Whether a plugin uses this is up to each plugin.
Changed
Section titled “Changed”- Less background work while you play. Stellar used to re-check your equipment, talents and dungeon death count on a timer every moment you were in the world; it now waits for the game to tell it something changed. Same information, noticeably less work per frame.
- Swapping a Battle Imagine now takes effect straight away. Before, Stellar kept showing the pair you had equipped when you logged in.
- Changing a piece of gear with the Replace button is picked up again — it used to go completely unnoticed.
- The dungeon death counter no longer gets stuck. If a later run reached the same number of deaths as an earlier one, it used to report none at all.
- Logging out and back in on another character no longer shows the previous character’s Psychoscope.
- Fixed a case where the game sent several changes at once and everything after the first one was thrown away.
Developer notes
Section titled “Developer notes”ILoadout.LiveState(LiveLoadoutState): the local player’s LIVE class + talents, parsed from the live line — never a saved plan. Refreshed with the loadout data; a respec re-fires the refresh through the new dirty-delta trigger.ILoadout.LiveStateChanged: ONE game-tick event for the whole build — equipped gear/module slots, class, talent stage/nodes, the equipped Battle Imagine pair, andIDeepSlumber.GetState(). Raised only after a re-read actually CHANGED what the service serves (structural, order-insensitive compares; a not-yet-read surface is no-signal). Published from the RESOLVE step, not the raw slot read, so it can never fire whileGetSlots()still describes the previous setup; an unresolvable change is held and delivered on the tick the data lands (LATE, never STALE). Raised on the game Update thread, unlikeIInventory.SelfGearChanged. Deep-Slumber joined this event in8e5a7b2(owner staging runsea/dXkw1PSyOG: a psychoscope factor unequipped between two archives and re-equipped after — the framework re-read it correctly but told nobody, so the consumer kept one stale snapshot across two materially different builds).IPluginServices.DeepSlumber(IDeepSlumber) +DeepSlumberState/DeepSlumberLine/DeepSlumberArea: live reflection reader overCharSerialize.SeasonCultivateLineData/SeasonRoleLevelData, seeded through the tolua# bridge because the C# mirror populates lazily and is a stale latch on a fresh session.zcontainermaps are iterated the game’s way —__pairsyields nil VALUES, so the reader indexes per key (owner runO1jJepsgKC). Session state clears on logout.IInventory.SelfGearChangedwidened to the self BUILD-state signal: genericContainerDirtyDeltaReader.TouchesField+ semantic wrappers fire it on talent (professionList, field 61) and Deep-Slumber (seasonCultivateLineData, field 101) method-22 deltas. Still network-thread — flag there, read on the tick, or subscribe toLiveStateChangedinstead.IResonanceState.Installedkeeps its shape but changes SOURCE and id space: equipped Battle Imagines are read from the skill hotbar’s aoyi slots 7/8 as aoyi SKILL ids (IGameDataResonance.GetImagineForSkillresolves them).CharSerialize.resonance(wire field 28) is never re-serialized on an in-session swap, so the old field-28 source could not see a swap at all (owner runsea/pNhmVQvVmV).- Wire fix: the top-level delta scan died after the first SKIPPED field because it never consumed that field’s trailing END tag — every change after the first in one packet was dropped.
- The framework’s last two per-tick game reads are gone. Live-state capture is driven by the container-merge event (replacing a field allowlist + a 1 s poll), and
IDungeonState.LastDefeatedCountnow ridesWorldNtf 3 EnterScene(EnterSceneInfo.SceneAttrs, the seed) +WorldNtf 7 SyncSceneAttrsinstead of a four-layer IL2CPP reflection read ofZWorld.GetWorldLuaAttr(348)on every main-thread beat — the same carrier the game’s own dungeon HUD watches. It also fixes a latent probe-side memo that never reset across runs (a second run reaching the same count reported 0 forever).StubRouter.Registeris now MULTICAST with registration order preserved (WorldNtf 3 has two subscribers and the order is load-bearing). BarElement.Overlay01/OverlayColor/OverlayInFront(BarStyle.Modernonly, #66): an optional SECOND fill fraction 0..1 on the SAME track asFraction01, re-pulled each refresh through its ownBarBinding.OverlayInFront: truedraws it over the main fill as a translucent band (the label overlay texts still sit on top);false(default)SetAsFirstSiblings it behind the opaque main fill so only the excess shows as an extension cap (e.g.Overlay01 = (hp+shield)/max). A nullOverlay01renders the bar exactly as before, andBarStyle.Defaultignores all three. Init-props only — additive, binary-compatible.tools/install-stellar.shgainsSTELLAR_FRAMEWORK_ONLY=1(deploy the framework set, plugin slots untouched) andSTELLAR_ONLY_PLUGINS=<slots>(deploy only the named plugin slots, framework slot untouched).
[2.1.0] - 2026-08-18
Section titled “[2.1.0] - 2026-08-18”2.1.0 (minor) — Stellar’s own menus now speak your language. Adds a plugin localization API; additive, binary-compatible with all existing plugins.
- Stellar’s settings menus now display in English, 日本語 (Japanese), ไทย (Thai), and Bahasa Indonesia. Pick your language in Settings → Themes → Language, or leave it on “Follow game client” to match your game.
Developer notes
Section titled “Developer notes”IPluginServices.Localization(ILocalization): a plugin-scoped UI-text localizer. Ship four embeddedLang/{en,ja,th,id}.jsoncatalogs (<EmbeddedResource Include="Lang/*.json" LogicalName="Lang.%(Filename)%(Extension)" />) and callLocalization.T("key")/TFormat("key", args); resolution is active-language → English → the key literal (a missing key renders visibly as the key). The framework auto-discovers each plugin’s catalogs at plugin-load (namespaced by plugin GUID, matched by theLang.<code>.jsonsuffix) — no registration code.ILocalizationControl(Settings-facing, NOT onIPluginServices) drives the settinglocalization.language(defaultfollow);ClientLanguageProbemaps the game client’sLanguageType(en=1,ja=2,th=5,id=6, elseen) to a supported code, and the setting live-switches (labels re-poll, baked renderers flush viaLanguageChanged). Catalog completeness is validated bytools/i18n-catalog.py <repo>(used/undefined/incomplete/orphan,--seed). Seedocs/plugin-development.md§ Localizing your plugin. New service only — additive, binary-compatible with plugins built against ≤2.0.3.- Section headers now stand out by accent colour (the theme’s
MenuAccent) instead of bold weight — crisp and clearly a header in every language. Why: under Proton/IL2CPP the overlay can’t render a readable bold for Thai/CJK — a real bold FACE will NOT load (Font.CreateDynamicFontFromOSFontverified across a host bold family name and a distinct normal-weight family installed both in the Wine prefix and as a hostfc-listfont — none render), and Unity’s synthetic bold blurs their tight loops/counters. So emphasis is carried by colour (WindowBuildersets emphasis text toMenuAccent;RegisterTextReskinre-applies it on a theme change), with weight kept CRISP — Latin still gets realFontStyle.Bold(it survives synthetic bold cleanly), complex scripts (CJK/Thai) stay regular. Titles take the crisp weight but no accent (already prominent by position).GlyphScriptpicks Latin-bold vs complex-regular per string, so it tracks a live language switch.
[2.0.3] - 2026-08-16
Section titled “[2.0.3] - 2026-08-16”2.0.3 (minor) — readable text inputs, plus internal groundwork for the CombatMeter Discord run-card and a reconnect fix that keeps a mid-dungeon disconnect from splitting your run in two.
- Text boxes in Stellar windows are readable again — they show the dark themed background instead of white-on-white.
Developer notes
Section titled “Developer notes”IGameAssets.LoadByPath— load an arbitrary game asset by its ZResLoader address (#61); backs the in-game image rendering for the CombatMeter Discord run-card.- Mid-dungeon reconnect party recovery:
PandaTeamInfoRefreshProbeinvokes the game’s ownWorldProxy.GetTeamInfo({})via the tolua# Lua bridge when in a dungeon withIPartySnapshot.PartyId == 0(a reconnect drops the party id until the game lazily re-delivers it — measured arriving only after a whole run). Bounded (≤3/run, 2.5 s throttle, world-gated main-thread); the reply is decoded by the existingPandaPartyStubProbe→PartyId. Enables CombatMeter’s reconnect run-split fix by making the party id available during the run instead of after it.
[2.0.2] - 2026-08-15
Section titled “[2.0.2] - 2026-08-15”- Plugins can now show more of your character’s stats — this powers the new Illusion-Breaking Strength readout in CombatMeter.
[2.0.1] - 2026-08-14
Section titled “[2.0.1] - 2026-08-14”2.0.1 (patch) — summoned-companion (imagine) damage now resolves to the right creature in combat data. Data-resolution only; no API change, binary-compatible with all existing plugins.
- Damage from your summoned companions now shows up correctly in the combat meter and on the logs website. Some newer companion abilities weren’t being recognised, so their damage went uncredited.
Developer notes
Section titled “Developer notes”GetImagineForSkillgainsImagineAoyiRule. Newer battle imagines flag their damage skills as SlotPositionId [0]/[6] (not the aoyi 7/8 slots) with noSkillFightLevelTablerow, so resolution rejected them and player-attributed summon damage never mapped back to the imagine (measured on jp/RXALtMH6J3: Celestial Flier 1008440, Rorola 2900840, Venobzzar 1007741, Kartgriff 111069 — all invisible while their damage sat in the run’s own perActorSkills). The new rung is gated to ids with no fight-level row of their own (baseId == 0): it decomposesMonsterId*100+NNover a lazySkillAoyiTableMonsterId→aoyi index (GameDataResonance.Aoyi.cs) plus a curated companion-arcane map (Boyce/Rorola/Fafala, table-evidenced). The gate is load-bearing — leveled player ids share the numeric namespace (140116 = Windborne Grace lv16 AND an Igoreus monster skill), and an own fight-level row wins. Negative memoisation stays pre-load-safe. Documented residuals: Igoreus/Denvel bands are ambiguous by construction; Dorothy/Lucy/Natsu have no evidenced rows and are never guessed. Also documents theNN=00composite probe as a sanctioned CombatMeter consumer contract (summon-entity appear-sourced imagine capture). Pinned byImagineAoyiRuleTests(23 cases incl. the collision band).
[2.0.0] - 2026-08-10
Section titled “[2.0.0] - 2026-08-10”2.0.0 (major) — the interface overhaul: resize the whole mod UI, a redesigned move-and-resize editing mode, a hotkey to hide the on-screen displays, and the Stellar menu on the login screen. One rendering engine now draws every window and overlay. Breaking for plugins that draw their own on-screen display (HUD) — the old HUD API is removed and those plugins must be rebuilt against the 2.0 SDK; all other plugins stay binary-compatible. Also bundles the 1.17.0–1.18.1 player fixes for anyone updating from an older build.
- A UI Scale slider in Settings → Themes that resizes the entire mod interface from 75% up to 150% (5% steps, default 100%). Every window and on-screen display scales together, and it previews live as you drag. Useful on very high-resolution screens where the panels felt small, or just to make everything bigger or more compact to taste.
- The mod interface now sizes itself to your screen resolution, so it stays proportional on any display instead of being tied to raw pixels.
- A redesigned layout-editing mode for moving and resizing the mod’s windows and on-screen displays. Turn it on with Alt+E, or with the new “Enter layout editing” button in Settings → Game UI; a bar across the top with an Exit button shows you’re editing. While editing: every on-screen display is shown — even ones you normally keep hidden — so you can position them; any window you can resize shows a small square in its bottom-right corner you can drag; and dragging stays smooth at any UI scale. When you leave edit mode, windows lock in place again.
- A hotkey to hide or show all of the mod’s on-screen displays at once, set to Alt+H by default (this resets each time you restart the game). There’s also an optional hold-to-hide key, unbound by default, that hides them only while you hold it — handy for a quick clean screenshot.
- The mod’s on-screen displays now automatically get out of the way while a game confirmation or OK pop-up is open, so they never sit on top of it.
- You can now open the Stellar menu right from the game’s login screen — a Stellar button sits in the login side bar, matching the in-world one — and the mod’s own windows now work at the title and login screens too, not only once you’re in the world.
- Settings now show each plugin’s proper name instead of its internal file name, and the Hotkeys screen groups shortcuts under the plugin that owns them, each with a short description of what it does.
Changed
Section titled “Changed”- This is a major update to how the mod draws its interface. Most plugins are unaffected, but a plugin that adds its own on-screen display may need an updated version from its author to work with 2.0.
- What this means for your existing setup: your saved window and on-screen-display positions are kept, and are remembered separately for each screen resolution. Because the interface now scales to your resolution, the windows may look a little smaller or larger than they did in the old version — most noticeable if you don’t play at 2560×1440 (smaller below that, larger above). Nothing is lost; use the new UI Scale slider and the layout-editing mode to set the size and positions you want.
- You now move and resize the mod’s windows only after entering layout-editing mode (Alt+E). Previously you could drag them at any time; locking them outside edit mode stops windows getting nudged by accident during play. Your positions are unchanged.
- The option that stops the game from also reacting to your mod shortcuts is now turned on by default. If you never changed this setting, the game will no longer also respond to a key that one of your mod hotkeys uses. If you had already set it yourself, your choice is kept exactly as it was.
- The default shortcut for layout-editing mode is now Alt+E (it used to be Shift+backtick). This only changes things if you never rebound it — a custom binding is kept. The block-hotkeys option is also now clearly worded: “Stop the game from also reacting to these keys.”
- Plugins can now build taller, wider, or full-width bars, put a larger label centred on top of the bar, and use a flatter “meter” look with an optional soft moving shine — so a plugin can present a prominent bar like a target’s HP bar or match the Combat Meter’s style. Whether a plugin uses this is up to each plugin.
- Your own class icon (the profession crest) now shows correctly.
- Switching game accounts no longer briefly shows the previous account’s character data — it’s cleared when you log out.
- Your window and on-screen-display positions now survive entering a dungeon or changing zones, instead of jumping back to their defaults.
- Class, skill, buff, and item icons that occasionally loaded blank now retry and appear reliably.
- An on-screen display you’ve hidden now stays hidden after you restart the game.
Developer notes
Section titled “Developer notes”- Breaking — the separate HUD toolkit is removed; there is now ONE UI path.
IHudHost/IHudHandle/HudSpec/HudAnchor,HudService(+.Layout),IHudRenderer/HudRenderer,HudElementBuilder(+.Meter), andHudBarAnimatorare deleted. On-screen overlays register through the window API:Windows.Register(new WindowRegistration(new WindowSpec(..., WindowCategory.HUD, WindowPanelStyle.Borderless){ Surface = SurfaceStyle.HudOverlay, Draggable = true, EditModeDragOnly = true, ShouldRender = ... }, root)). The sharedHudElementrecord tree is unchanged. Newenum SurfaceStyle { Menu, HudOverlay }(defaultMenu→ every existing window byte-identical);HudOverlayreproduces the old HUD look (always-shadowed twin text honouringFontSize/DynamicFontSize/ShadowDistance, the rounded 9-slice smoothed-fill bar, the transparent pill) by reusingHudThemeAssets, now owned byWindowRenderer.IThemeHudColorsretained (Toast + the HudOverlay bake still use it). First-party PlayerHUD / RaidManager countdown / DebugInfo migrated in lockstep, pixel-identical. Only plugins that calledIPluginServices.Hud.Register(HudSpec)are affected; all others are source- and binary-compatible. - Source-breaking rename
BarStyle.Meter→BarStyle.Modern(enum value unchanged,Default=0/Modern=1 → binary-compatible for compiled plugins).BarElementgainsinit-only geometry (Height/Width/FillWidth/LabelFontSize/LabelInside) plus the opt-inModernmeter style with optionalSheen/SecondaryLabel; the original positional constructor is untouched (source- and binary-compatible, unset fields default to the old render byte-for-byte). Both the HUD-overlay path and the window path (WindowBuilder.BuildBar/WindowBuilder.PreviewBuildBarModernWindow) honour every new field. - UI scaling. Window overlay canvas gets a
CanvasScaler(ScaleWithScreenSize, reference 2560×1440, match 0.5). At the reference resolution scaleFactor = 1.0 (byte-identical to pre-2.0); other resolutions convert drag/resize/on-screen-clamp/dropdown-placement/click-rect between screen px and canvas units. The UI Scale slider setsreferenceResolution = (2560/u, 1440/u), persisted astheme.uiscale, quantised to 5% so the font atlas repacks only at boundaries;canvas.pixelPerfecttoggles off at fractional scaleFactor to kill dragged-window jitter.LayoutStorage.Getreads saved rects as canvas units (scaleFactor-corrected) while the KEY lookup stays raw, so pre-2.0 per-resolution layouts are still found. - Game-phase model.
GamePhase { Startup, TitleScreen, CharSelect, World }and[Flags] GameUIState(+GameHudHidden/AnyMenu/Blocking/Popuppresets) onIClientState(Phase/PhaseChanged/IsWorldActive/UiState). The framework tick now runs UI/input/hotkeys every phase (so plugin windows work at title/login); game-state work self-gates per-unit onIsWorldActive, enforced by new STELLAR0006 analyzer (a[WorldGated]method must early-return on!IsWorldActive). Window visibility is plugin-owned via requiredWindowSpec.ShouldRender(IRenderGated), replacingAutoHideBehindGameMenus/HideUntilInWorld; the gate fails safe (null/throwing predicate → hidden + warned) so a stale pre-2.0 plugin can’t NRE the tick.Loading/Popupare driven by dedicated un-gated probes (reliable across loading screens / confirm dialogstips_common_popup/tips_sys_dialog). - Login-screen integration.
NativeUiAnchor.LoginSidebarinjects the Stellar rail icon intologin_main’s sidebar; uGUI injection is un-gated with per-anchor phase-relevance (LoginSidebar=TitleScreen, MainMenuRail/HudTopRight=World) so there is no out-of-phase probing cost.LauncherEntry.ShouldShowgates individual launcher tiles by phase; framework chrome hides over the loading screen. - New members (all additive/binary-compatible):
GameTextureElement.CornerRadius,CooldownTileElement.FallbackLabel,IWindowHost.IsLayoutEditing. - Plugin identity/hotkeys. The framework adopts each plugin’s declared
IStellarPlugin.Nameas its display name (Plugins panel, per-plugin Performance rows, Hotkeys group header); hotkeys group by owning plugin via an internal owner-tagged declaration sink, andIHotkeyActiongainsPluginId+Description. HUD-visibility adds internalIInputGateway.IsKeyHeld+HotkeyService.IsActionHeld. - Interop foundation (toward v2.1), additive, framework-only-implemented:
StellarInterop(cached reflection floor),ILua(IPluginServices.Lua),IHarmonyHost(IPluginServices.Harmony, auto-unpatch on plugin dispose), andIFramework/IFrameworkTiming(Post/Every/TimeNow, skew-correctedICombatSnapshot.ServerNow). - Icon-load resilience. Failed icon slots retry with bounded backoff (~0.5–2s, up to 4×) instead of memoizing
Failedfor the session; a first request that beats the HybridCLR loader registration self-heals. Fixes blank profession crest / skill / buff / item icons. - Other fixes: profession-crest projection falls back to
ProfessionSystemTable.Iconand guards against table-not-ready; account/character data clears on logout (account-switch leak); mod-window and native-game-HUD positions survive scene change; a persisted layout-editor hide survives relaunch. Throwaway phase-diag overlay removed. Merged from main: Steam client region detection (1.18.1) and loadout talent-spec correctness (1.18.0). - Test suite: 1012 passing.
[1.18.1] - 2026-08-07
Section titled “[1.18.1] - 2026-08-07”1.18.1 (patch) — Steam client region detection fix so runs from the Steam build upload again. Detection-logic only; no API change, binary-compatible with all existing plugins.
- If you play the game through Steam, your dungeon runs now upload to the logs website again. The mod couldn’t recognise the Steam version of the game and was holding those runs back.
Developer notes
Section titled “Developer notes”- GameEnvironmentService now matches the running executable by name PREFIX (StarSEA → SEA, StarASIA → JP) instead of exact filename, so the Steam build’s StarSEA_STEAM.exe resolves to SEA instead of Unknown — Unknown made CombatMeter’s RegionKnownOrWarn withhold every upload. One row per region covers StarLauncher, Steam, and future store builds. Regression test SteamSeaExecutable_DetectsSea; the existing UnknownExecutable_DetectsUnknown still pins non-matches. The Steam install dir is flat (no release_<ver> segment) so GameVersion reads “unknown” there — cosmetic, not gated on for uploads. Commit 5f976ea.
[1.18.0] - 2026-08-06
Section titled “[1.18.0] - 2026-08-06”1.18.0 (minor) — one additive loadout fix on top of 1.17.0: the talent specialization shown for a saved build is now read consistently with that build’s own talent nodes. Binary-compatible with plugins built against ≤1.16.1.
- Your saved talent build now shows the correct specialization on the logs website, instead of sometimes showing a different one than the talents you actually picked.
Developer notes
Section titled “Developer notes”- PandaLoadoutProbe.Resolution now sources a saved plan’s talent stage from the live per-profession container (talentList[prof].talentStageCfgId) so the stage id agrees with the plan’s own talent nodes — fixes a stale talentStageCfgId that mislabeled the spec (e.g. run sea/ZEEJjddKHN). Single-file change to PandaLoadoutProbe.Resolution.cs; cherry-picked from e73b99b. The exploratory NotifyShowTips (m42) win-tip tap from the same session was dropped as inert — vaults do send a real settlement message, so the actual vault fix lives plugin-side (CombatMeter 1.6.1). The SquareHandle slider knob referenced in FrameworkVersion already shipped in the 1.17.0 tree and is unchanged here.
[1.17.0] - 2026-08-05
Section titled “[1.17.0] - 2026-08-05”1.17.0 (minor) — kill detection from the game’s own logic, a run-id fix for instanced dungeons, per-class equipped-loadout capture, and a mounted-state stats-survival fix. Additive; binary-compatible with plugins built against ≤1.16.1.
- The Combat Meter can now tell exactly when an enemy dies — including bosses finished off by a cutscene or scripted move, not just a normal killing blow — so fights are timed and scored correctly.
- The logs website can now show the gear, modules, and talents you currently have equipped, for each class you play.
- Some dungeon runs weren’t being saved to the logs website — they’re recorded correctly now.
- Your character’s stats and info no longer disappear while you’re mounted.
Developer notes
Section titled “Developer notes”- Run-id: 3.7 instanced/Mistveil scenes whose scene uuid is below 2^53 were misclassified by magnitude; now classified by SceneType and early at the wire via AttrSceneBasicId (341). Fixes “No run id” (ranked runs not uploading). Commits 4683d8d, 9cb72c8.
- LIVE per-class loadout: parse the refresh chunk’s LIVE row (
_StellarLiveProbe) so a played class exposes its currently-equipped gear/modules/talents, and it is the sole source when a class has no saved plan. Commit ba74dd3 (+ the per-class gear/modules/talents series a513b29..c1a9d65). - Player-state: rescue the local player entity when it goes dark while mounted so PlayerStats/identity survive mounting; identity served from the char record, not the world entity. Commits 1107646, 615042d, 7a40e28.
- New
Stellar.Abstractions.Domain.ActorStateenum (Dead=9,Breaking=23,Unknown=0 for any other/future wire value) andCombatEvent.EntityStateChanged(TimestampMs, TargetId, ActorState), riding the existingICombatEventsstream — no service interface gains a member, so STELLAR0005’s 8-member ceiling is untouched. Field-proven after three recon rounds — seerecon/entity-state-death-signal-notes.md. Round 1’s spec’d leaf,Panda.ZGame.EntityCtrlDead.OnEnter, installed cleanly but stayed silent across all ten deaths in the owner’s confirming run: disproven. Round 2 tried the widerPanda.ZGame.ZStateMachine.onStateChanged/EnterStatehooks on top;onStateChangednever fired (a real negative result), butEnterStatefired correctly for every one of the ten deaths and resolvedDeadcorrectly — it was not broken, just costlier (it fires on every actor’s every transition, not just the one we care about), so round 3 drops it in favour of the cheaper leaf and keeps it on record as a field-proven fallback ifZStateDeadis ever removed/renamed. The shipped design patches exactly two sites:Panda.ZGame.ZStateDead.OnEnter(PROVEN — all ten deaths) andPanda.ZGame.ZStateBreaking.OnEnter(untested, not disproven; kept as the direct sibling of the proven hook, and as a timestamp source for an open frametime-spike investigation). Each site resolves/installs independently and degrades to “signal off” (logged) rather than throwing if a type/accessor is missing after a future game patch — seePandaEntityStateProbe. The site-attributed, per-scene-budgeted ungated observation line survives the cut-down unchanged. 2026-07-28 entity-state-death-signal spec; retires the HP-inference gone-timeout forArchiveReason.BossKillon the plugin side (follow-up work, not in this release).
[1.16.1] - 2026-07-25
Section titled “[1.16.1] - 2026-07-25”1.16.1 — same code as 1.16.0, re-cut under a fresh bundle filename because a CDN cache mismatch left some 1.16.0 downloads stuck at 100%. Carries the full 1.16.0 patch notes so players updating straight from 1.15.0 see what changed.
- Update download no longer gets stuck at 100% for some players (a caching problem on our download server with the 1.16.0 file — this version uses a fresh file).
- Much higher FPS in dungeons and busy areas. The framework was quietly doing heavy background work several times per second — on our test machine that alone cost up to 50 FPS in a dungeon (86 to 144 after the fix). This work is now nearly free.
- The regular micro-stutter is gone. If your frametime graph showed small spikes about 5 times every second — even with no plugins installed — that was us. Fixed.
- No more short freezes every few seconds. The framework re-checked your whole inventory once per second and threw the result away, which piled up memory and caused brief frozen frames (worst on lower-end PCs). It now only does that work when your inventory actually changes.
- Smoother big fights. Network and combat data is now processed with far less memory churn, so crowded areas and boss fights cause fewer hitches.
- Zero input overhead unless you use key-blocking. The hotkey “block from game” feature used to sit on the game’s keyboard input all the time; now it activates only while you actually have a blocked hotkey or are recording a new one.
Developer notes
Section titled “Developer notes”- Identical framework code to 1.16.0 (only FrameworkVersion + this changelog differ). The 1.16.0 manifest republish (player-facing patch notes, PR #46) rotated the bundle sha256 while CDN edges still held the previous zip (max-age 14400); launchers on those edges failed the hash check and hung at 100%. A fresh
Stellar-1.16.1.zipfilename has no cached copies anywhere, so every edge serves consistent manifest+bundle. Lesson (now in docs/release-process.md): a same-version republish requires an immediate CDN purge; prefer a patch bump instead. - Patch-note bullets are now plain text (the launcher renders no markdown — 1.16.0’s
**showed literally).
[1.16.0] - 2026-07-25
Section titled “[1.16.0] - 2026-07-25”1.16.0 (minor) — the frametime-jitter release: eliminates the framework’s in-dungeon FPS loss and frame-spike comb, root-caused by matched A/B on a live client (PR #44). Additive; binary-compatible with plugins built against ≤1.15.0.
- Much higher FPS in dungeons and busy areas. The framework was quietly doing heavy background work several times per second — on our test machine that alone cost up to 50 FPS in a dungeon (86 → 144 after the fix). This work is now nearly free.
- The regular micro-stutter is gone. If your frametime graph showed small spikes about 5 times every second — even with no plugins installed — that was us. Fixed.
- No more short freezes every few seconds. The framework re-checked your whole inventory once per second and threw the result away, which piled up memory and caused brief frozen frames (worst on lower-end PCs). It now only does that work when your inventory actually changes.
- Smoother big fights. Network and combat data is now processed with far less memory churn, so crowded areas and boss fights cause fewer hitches.
- Zero input overhead unless you use key-blocking. The hotkey “block from game” feature used to sit on the game’s keyboard input all the time; now it activates only while you actually have a blocked hotkey or are recording a new one.
Developer notes
Section titled “Developer notes”These details never appear in the launcher — full technical background lives here and in PR #44.
- Anchor probe:
PandaUGuiAdapterran a path-formGameObject.Find(full scene walk, ~30 ms/hit in dense scenes) every 200 ms while the game menu was closed; now cachedzuiroot+ relativeTransform.Find, active-only contract preserved. A check-standards blocker now bans path-formGameObject.Findliterals (exemptions: devkit tech-debt D-26). - Inventory: the 1 Hz poll rebuilt the full module inventory via reflection (~1.8 MB/s garbage → 100–475 ms stop-the-world GC frames on low-spec); now generation-guarded on actual inventory syncs. Measured avgUpdAlloc 55 → 9.8 KB/tick.
- Wire tap: handler snapshot resolved before payload materialization (unsubscribed AOI/world traffic no longer copied), zero-copy reassembly drain, cached per-thread zstd decompressor with exact-size output, quiet-connection eviction + drained-buffer shrink.
- Input:
Rewired.Keyboard.GetKey*prefixes (~8–25k calls/s at render rate) install on first block/capture and uninstall when cleared; primary-key pre-filter before modifier interop reads. - Combat parse: AOI attr payloads slice the per-packet array (−45% alloc per combat packet), lazy cooldown list, positions cache uses an Interlocked counter instead of per-update
ConcurrentDictionary.Count, compiled delegate replacesMethodInfo.Invokefor the IL2CPP span extractor. - New PERFHUD diagnostics:
PerfProbe.HookEndRewired+hook:rewired=ms/callsin the[Perf]line, andBeginSegcoverage for the previously unsegmented tick items.
[1.15.0] - 2026-07-21
Section titled “[1.15.0] - 2026-07-21”1.15.0 (minor) — adds per-plugin binary file storage, the substrate for CombatMeter’s byte-for-byte re-upload. Additive; binary-compatible with plugins built against ≤1.14.0.
IPluginServices.Data(IPluginDataStore) — per-plugin binary file storage (Write/Read/Delete/List) for data too large or opaque forIConfigSection. Never-throws; names are path-traversal-safe (rejects.., rooted paths, backslashes,>1separator). Each plugin’s store is rooted at<gameRoot>/stellar/plugindata/<guid>.data/— a sibling of, and deliberately OUTSIDE, the recursivestellar/plugins/DLL scan path (FrameworkPaths, with a pinned non-nesting test), so a plugin’s stored files can never be shadow-loaded as assemblies.
[1.14.0] - 2026-07-18
Section titled “[1.14.0] - 2026-07-18”1.14.0 (minor) — CombatMeter sync-fix surface: honest death inference, live party status transport, and dungeon flow-state. Additive; binary-compatible with plugins built against ≤1.13.0.
EntityVitals.HasHpObservation— true only once a real current-HP value has been observed; a MaxHp-only attr delta now reads “alive, HP unknown” instead of dead. AOI appear packets now seed vitals directly (previously the first delta after appear defined them).PartyMember.FastSyncState— rawTeamMemberFastSyncData.state(field 6), previously parsed and dropped; the game client itself ignores this field, so semantics are calibrated empirically and consumers must treat unmapped values as “no signal”.IDungeonState.CurrentFlowState+FlowStateVersion— the dungeon flow state machine (EDungeonState: Active/Ready/Playing/End/Settlement/Vote) surfaced from both the method-23 full sync and the method-24 dirty delta, with a monotonic transition counter as the poll-friendly change notification.
- Party fast-sync deliveries that change ONLY the member’s status field now fire
MemberUpdated.
[1.13.0] - 2026-07-16
Section titled “[1.13.0] - 2026-07-16”1.13.0 (minor) — adds live Trading-Center membership so exchange plugins categorize new items without a rebuild. Additive; binary-compatible with plugins built against ≤1.12.0.
IExchange.GetStallSubcategoryMap()— item config id → Trading-Center subcategory leaf (101-104 Growth / 201-209 Life Skills / 301 Modules / 401-405 Appearance), read live at runtime from the game’sStallDetailTable. Empty when the table isn’t loaded yet; consumers fall back to their own data. Lets the Exchange Buyer plugin surface Season-3 (and future) items with no rebuild.
- Game config tables keyed by id in the
ZTable<int, row>key (rather than a value column) — e.g.StallDetailTable— are now read via a key-aware path; the previous value-only table loader silently returned 0 rows for them.
[1.12.0] - 2026-07-11
Section titled “[1.12.0] - 2026-07-11”1.12.0 (minor) — adds the game-environment identity service for SEA/JP region-aware uploads (spec: server-region partitioning). Additive; binary-compatible with plugins built against ≤1.11.0.
IGameEnvironment— region (GameRegionSEA/JP/Unknown, lowercaseRegionCodewire form) + installedGameVersion, detected once at boot from install markers (StarSEA.exe→ SEA); framework configenvironment.regionoverrides. Boot log prints[Stellar] region=… version=… source=….
[1.11.0] - 2026-07-10
Section titled “[1.11.0] - 2026-07-10”1.11.0 (minor) — consolidates the portraits/replay feature line onto main (merge ad173a0): every addition is a new service, init-only property, or defaulted parameter under the plugins-consume-never-implement contract, binary-compatible with plugins built against ≤1.10.0. Note: the IEntityTransforms / IGameDataWorld.GetMonsterByEntity entries listed under the 1.10.0 sections below were documented there but did not ship in the 1.10.0 bundle; they ship in this release.
IDungeonState— dungeon lifecycle + settlement service: run outcome (DungeonOutcome),DungeonSettlementInfo (PassTimeSeconds, MasterModeScore, TotalScore)with the achieved total-score carried end-to-end, defeated count, and run-identity gating for fail-outs.IEntityTransforms— live world transform (position + facing) reads of arbitrary entities by id (main-thread), for replay/position capture.IEntityDetail.RefreshSocialSnapshot— self social-data refresh (drives the game’sAsyncGetSocialDatavia Lua) so plugins can re-read the local player’s social snapshot (e.g. master score) after a settlement without waiting for cache expiry.SocialSnapshot.MasterScoreand related identity fields.IGameDataWorld.GetMonsterByEntity(EntityId)+MonsterInfo.MonsterType/IsBoss— entity→monster-table resolution through the shared combat entity tracker.- Skill-phase
CombatEventcases (cast phases 101–105) for forensic capture. GearInstance.BreakThroughTime;PartyMemberadditions;IPluginServicesexposes the new toolkit services.
[1.10.0] — 2026-07-03
Section titled “[1.10.0] — 2026-07-03”1.10.0 (minor) — pure interface additions under the plugins-consume-never-implement
contract (IEntityTransforms, IGameDataWorld.GetMonsterByEntity); MonsterInfo’s new MonsterType/
IsBoss are init-only properties, not primary-ctor params, so the type stays binary-compatible with
plugins built against ≤1.9.1 (see the SliderElement/1.7.1 precedent below).
IEntityTransforms— toolkit service reading the live world transform (position + facing) of an arbitrary entity by id, for replay/position capture.TryGetTransformreturnsfalse(leaving the out-params at their defaults) when the entity isn’t resolvable this frame; reads must happen on the main thread.IGameDataWorld.GetMonsterByEntity(EntityId)— resolves a live entity to itsMonsterInfotable row via the entity’s cached attr-10 (config-id) attribute, routed through a sharedCombatEntityTrackerso combat and game-data probes read the same cached attribute set instead of each maintaining their own.MonsterInfo.MonsterType/MonsterInfo.IsBoss— numeric monster classification (EMonsterType-mirrored; 0=Monster, 1=Elite, 2=Boss) and a derived boss flag, loaded from theMonsterTablerow. Confirmed by recon on the Ancient Purifier run: entity attr 10 →MonsterTable[33301].MonsterType == 2.
- Run-identity leak across non-instanced scenes.
IDungeonState.CurrentRunId(and the internalDungeonRunIdGate) now clears to 0 on entering a non-instanced (town/field) scene, instead of letting a previous dungeon run’s id linger and get attributed to unrelated open-world activity.
Removed
Section titled “Removed”- Recon scaffolding for boss identification (the entity → monster-config-id spike diagnostic and the
short-lived
IMonsterCatalog/MonsterCatalogServiceprobe path) — superseded by theGetMonsterByEntityMonsterInfo.MonsterType/IsBossresolution above and no longer needed.
[1.10.0] - 2026-07-08
Section titled “[1.10.0] - 2026-07-08”- Window-framework support behind the CooldownBar overlay (all additive + binary-compatible). (#32)
WindowSpec.BackgroundOpacity(Func<float>?) — poll-diffed black backdrop on the borderless root’s click-blocker Image; fills the full window rect and expands on height resize.ColumnElement.Padding(int) — uniform inner padding on all four sides.RowElement.Justify+RowJustifyenum (Left/Center/Right) — ships a 2-arg compat overload for plugins built against the old signature.BackdropElement— stretch-fillignoreLayoutbackdrop (place first in a column so siblings draw on top).VirtualListElement.ResetScroll(Func<bool>?) — polled each refresh; true snaps the list to the top.CooldownTileElement.OnClick(Action?) — whole-tile click action.
Changed
Section titled “Changed”- Atomic game-asset icons —
GameTextureElementmoved toWindowBuilder.GameTexture.cs; icon rebind is now aWindowTokenbinding applied in the sameApply()pass as the virtual list, removing the one-frame wrong-icon blink while scrolling. Buff icons resolve as atlasSprites instead ofTexture2D. ConditionalElementflex clamp — aCondnode inside a row no longer setsflexibleWidth=1(it was stealing all slack, centring fixed content in a borderless bar); column parents still force-expand.- VirtualList viewport inset 9px on the right so the scrollbar no longer overlaps content.
- Framework deploy target moved out of the csproj into
Local.props.
[1.9.1] - 2026-07-01
Section titled “[1.9.1] - 2026-07-01”AutoHideBehindGameMenuswindows now hide behind more full-screen game screens.PandaMenuStateProbewas extended to detect NPC dialogue, loading screens, the dungeon-enter confirm popup, the line-selector panel, and story cutscenes as full-screen menu states. (#29)- The dungeon-enter confirm view instantiates as
team_copy_popuponUILayerFuncPopup(notcommon_matchingonUILayerTop), so detection now scansUILayerFuncPopup(AnyChildActive) and checks themain_line_windowprefix onUILayerMain. - Removed the probe’s diagnostic-logging infrastructure (
Log,DumpZuiroot,DumpActiveCanvases); the layer map is documented inGameMenuState.mdinstead.
- The dungeon-enter confirm view instantiates as
[1.9.0] - 2026-06-30
Section titled “[1.9.0] - 2026-06-30”IWindowControl.BringToFront()— explicitly raises a window above others via aZFrontcounter that sorts above category (ZCat), so cross-category fronting works (e.g. a HUD window surfaces above a Tools window). ABringToFrontPendingflag covers the still-hidden case, soSetVisible(true)+BringToFront()in the same line always surfaces on top. (#27)
- Interaction pass-through to covered windows. A
FrontWindowBlocksguard now wraps every hit-test inWindowInteractionTicker(hovers, resize grip, drag areas, scrollbar suppression, chart pan/zoom, chart navigator, render-host drag/zoom), so back-window elements (chart, 3D portrait, scrollbar) no longer receive interaction when a front window covers the pointer. (#27) - Dropdown item styling now matches
SelectableElement: roundedSwatchBgchip, transparent rest / [email protected] hover / [email protected] active selection, VLG 6/4 padding,ContentSizeFitter. Hover is ticker-driven (Unity EventSystemColorTintdoesn’t fire reliably alongside the ticker). (#27)
[1.8.0] - 2026-06-28
Section titled “[1.8.0] - 2026-06-28”DropdownElement— a reusable compact dropdown for a small, fixed set of mutually-exclusive choices. The trigger shows the current option (caption + ▾); clicking it opens a themed floating option list that floats above the window’s scroll clip (parented to the canvas root, so a dropdown inside aScrollElementis not clipped by itsRectMask2D). Picking an option calls back with its index; the list dismisses on pick, outside-click (a full-screen invisible blocker), or Escape. Drop it into any window element tree like the other widgets.
Changed
Section titled “Changed”- The Settings → Performance per-plugin Self-rate control (Off / Boost / Self-managed) is now a
DropdownElementinstead of a click-to-cycle button.
[1.7.1] - 2026-06-28
Section titled “[1.7.1] - 2026-06-28”- Binary-compatibility regression from 1.7.0.
SliderElement’s newWidth/HandleSizewere added to the record’s primary constructor, which changed the generated constructor signature and broke the old positional ctor. Plugins compiled against framework ≤1.6.0 that construct aSliderElement(e.g.Stellar.AutoFishing) failed to load on 1.7.0 with aTargetInvocationExceptionwrappingMissingMethodException.Width/HandleSizeare now init-only properties (the original primary ctor is restored), so those plugins load again.
[1.7.0] - 2026-06-28
Section titled “[1.7.0] - 2026-06-28”- Per-plugin & dynamic update rate. Each plugin’s
IFramework.Updatenow ticks at its own rate instead of being welded to the single global rate. In Settings → Performance a new “Per-plugin update rate” section lets the user set each plugin’s rate (default Follow global) and grant a per-plugin Self-rate permission (Off / Boost / Self-managed). IFramework.RequestUpdateRate(int hz)→IUpdateRateScope— a permitted plugin can temporarily ramp its own tick rate (e.g. a market snipe’s fast-retry window) and revert by disposing the scope. Returns an inert no-op scope unless the user granted the plugin permission. “Self-managed” exempts the ramp from the 10 s leak-guard for plugins that hold an elevated rate indefinitely.SliderElement.Width/SliderElement.HandleSizefor compact, fixed-size sliders;ToggleElement.Enabledis now honored by the renderer; fixed-widthNoWraptext now clips to its column.
Changed
Section titled “Changed”- The framework tick is now a single variable-speed clock at
max(global, all active plugin rates)(clamped[10,240], realized ≤ frame rate). The Lua-bridge probe drains ride the master tick so a ramped plugin’s RPC round-trips complete proportionally faster; expensive draw/refresh work stays gated to the global rate. Idle behaviour (nothing ramped) is unchanged from 1.6.0.
- Perf overlay no longer freezes its per-window timings when the master rate exceeds the global rate.
[1.6.0] - 2026-06-25
Section titled “[1.6.0] - 2026-06-25”IExchange— player Trading-Center market access, exposed asIPluginServices.Market. Plugins can query the live market and buy through the game’s own trade flow — no packet construction; every purchase is built and validated server-side. Driven via the game’s WorldProxy exchange RPCs (“Approach A”, headless — the trade page never needs to be open).- Reads:
QueryListingsAsync(itemId)(live listings, cheapest-first),QueryCatalogAsync(category)(category browse — the request’stypeis theStallCategoryfamily, derived from the leaf id),QueryCareListAsync(kind)(watch list + availability),QueryNoticeAsync(itemId)(scheduled/pre-order listings). - Buy:
BuyAsync(itemId, quantity, price)→ExchangeBuyOutcome(Success/NoItemAvailable/InsufficientFunds/Rejected/Timeout). - DTOs:
ExchangeListing,ExchangeCatalogItem,ExchangeCareItem,ExchangeNoticeListing,ExchangeItemKind.
- Reads:
- Extensible toast notifications.
INotifications.Create()fluent builder (WithMessage/WithKind/WithDuration/WithIcon) with custom-icon support — render a supplied texture (e.g. a game item icon) in the toast’s icon slot;nullfalls back to the baked per-kind glyph. The existingNotify(string, kind, seconds)shortcut is unchanged.
[1.5.0] - 2026-06-24
Section titled “[1.5.0] - 2026-06-24”- Team voice & dungeon ready-check, surfaced as typed party events. The framework decodes the
game’s voice and ready-check wire traffic and exposes it through
IPartyEvents/IPartyRoster, so plugins consume clean events instead of touching IL2CPP or Lua. (#21)- Ready-check:
WorldNtfLuaStubDispatcher(HarmonyX postfix onZLuaStub.OnCallStub) catches methods 70/71, which flow through the Lua stub rather thanWorldNtfStub;NotifyReadyCheckReaderdecodesNotifyAllMemberReady(70) /NotifyCaptainReady(71) intoIPartyEvents.ReadyCheckResponded/ReadyCheckPhaseChanged. - Team voice:
GrpcTeamNtfmethods 25/26 (mic mode / speaking) decoded inNotifyTeamVoiceReaders;voice_is_open(TeamMemDataf7) +mem_real_time_voice_infos(GetTeamInfoReplyf4) parsed for correct state on join/relogin (incl. theOpenSpeakeredge case the bool can’t express). NewMicrophoneStatusenum +IPartyRoster.GetMicStatus/IsSpeaking— additive;PartyMemberstays binary-compatible. - Meter row (
MeterRowData):NameColor(ready-check vote tint),VoiceIcon/VoiceIconTint/ShowVoiceIcon(mic icon),RowBorder(green while talking),CrestTint.
- Ready-check:
- UI primitives for click-away popups.
WindowSpec.DismissOnOutsideClick— Escape or press-outside invokesOnCloseand hides the window;IsShownstays in sync.PanelElement— themed popup container (2 px border + lifted background + padded content host).- Deterministic z-ordering (
ReorderWindows): draw order follows(ZOrder, Category, Id)regardless of plugin mount order, so click-away popups always render on top.
- Meter row border now draws all four edges (previously only the top edge was visible).
WindowRenderer.SetRectclamps programmatic resizes toMinWidth/MaxWidth. Previously only drag-resize was clamped, so mode switches,RefreshPartyFocusHeight, and prefs restore could silently push a window below its registered minimum.
[1.4.2] - 2026-06-22
Section titled “[1.4.2] - 2026-06-22”- Non-ASCII text input (e.g. Thai) truncated in uGUI fields. Switched from
onValidateInput(per-char ASCII gate) toonValueChanged, so multi-byte / IME input is preserved. (#19)
[1.4.1] - 2026-06-22
Section titled “[1.4.1] - 2026-06-22”- Periodic ~200 ms in-game freeze from the rail-button template lookup. The lookup walked the entire UI tree on a timer; it’s now scoped to the menu-panel subtree. (#17)
[1.4.0] - 2026-06-22
Section titled “[1.4.0] - 2026-06-22”IWindowControl.SetVisiblePersist(bool)— show/hide a window AND persist the choice to the active layout slot (per resolution), so it survives relaunch. This is the single source of truth the framework reapplies on launch (the layout-editor eye-toggle writes the same slot). Plugins should use it for user-driven visibility toggles (hotkeys, close buttons) instead ofSetVisibleplus a private config key, which desyncs from the slot and loses to it on relaunch.SetVisibleis now documented as session-only (non-persisting).
- Native-UI edit-mode grab-boxes dropped to the bottom-left corner during loading / cutscenes. The game collapses its HUD to ~1px stubs during those transitions and the edit outline followed them down. It now holds each element’s last real-size rect (carried across the scene-change re-resolve) and never caches a collapsed stub.
- Repositioned game-UI elements flung off-screen / left at the game default after a cutscene.
SetRectno longer runs on an inactive element (its world-corners are garbage mid-cutscene), caps any bogus translate, and re-applies the saved position once the game resets the element; the 1 Hz re-assert no longer force-shows elements the game is hiding for a cutscene.
[1.3.0] - 2026-06-21
Section titled “[1.3.0] - 2026-06-21”- Native notice banners (
INoticeTipsonIPluginServices) — trigger the game’s own notice system (dungeon bars, win/fail banners, pop tips) with full control over content and audio via a fluent builder (Create/WithContent/WithAudio/WithDuration/Show).NoticeTipService.Show()is thread-safe: it enqueues the pre-built Lua chunk to aConcurrentQueuedrained on the main-thread tick, so plugins may call it from async continuations or any thread without marshalling boilerplate. (#15) - Consume hotkey keypresses (Settings → Hotkeys) — a global toggle that blocks bound keys from
reaching the game via Rewired while the framework still receives them through
UnityEngine.Input. Blocking is modifier-aware (Ctrl+F1 bound does not also block bare F1), and all keys are suppressed while a rebind cell is open. Backed byIHotkeyBlockDirectory+HotkeyKeyBlockPatch. (#15)
[1.2.0] - 2026-06-19
Section titled “[1.2.0] - 2026-06-19”- Loadout API (
ILoadoutonIPluginServices) — read the player’s saved in-game loadouts (the game’s “Role Plan” system: class + gear + spec + modules) and switch to one. Drives the game’s ownAsyncSwitchRolePlan(the path the in-game dropdown uses), so all server-side validation (combat lock, etc.) is respected, never bypassed. Backed byPandaLoadoutProbeover the tolua# Lua bridge +WorldProxy. - Notification toasts (
INotificationsonIPluginServices) — transient on-screen toasts any plugin can raise (Notify(message, kind, seconds)), rendered top-centre with a Pop+Scale animation, per-kind colour, and a countdown bar. First consumer: the LoadoutSwitcher plugin.
[1.1.2] - 2026-06-18
Section titled “[1.1.2] - 2026-06-18”- Launcher tile icons are now live-bound: a plugin icon that loads after the launcher is built (plugins register asynchronously, and the Full / Minimal layouts materialise their tiles at different times) no longer stays baked to the generic puzzle-piece fallback — it now appears in every launcher mode. Fixes a plugin icon showing in expanded mode but the puzzle-piece in minimal mode. (#11)
- Game-data name resolution: locale-gate the empty-
Name→NameDesignfallback so the design-name fallback only kicks in where it should, instead of leaking into clients that do have a localized name. (#12)
[1.1.1] - 2026-06-18
Section titled “[1.1.1] - 2026-06-18”- Plugin SDK on NuGet.org —
Stellar.Abstractions,Stellar.PluginContracts, andStellar.Plugin.InteropRefs(the Unity/Il2Cpp/BepInEx compile-time reference stubs) are published via Trusted Publishing, so plugins build with just<PackageReference>s — no framework checkout and no game install. First step toward per-plugin repos (see the DevKit’s DIP17 migration plan).
Changed
Section titled “Changed”- Framework runtime is unchanged from 1.1.0 (this release adds the SDK + contributor docs only).
[1.1.0] - 2026-06-18
Section titled “[1.1.0] - 2026-06-18”LineChartElement— multi-series time-series line chart with labelled X/Y axes, axis titles, a legend, auto-scaled Y, and interactive zoom (scroll/drag + −/+/Reset buttons + range scrollbar). Rendered via an injectedMaskableGraphicmesh (ChartGraphic). First consumer: CombatMeter history charts.- Generic inter-plugin exchange:
IPluginExchange(Provide<T>/Consume<T>) onIPluginServices, brokered purely byType, plus the newStellar.PluginContractsassembly that cooperating plugins reference (alongsideStellar.Abstractions) for shared contracts —FrozenEntityandIFrozenEntityViewer. (#5 Phase 2) - HUD text enhancements on
TextElement:FontSize,DynamicFontSize,ShadowDistance, andNoWrap, plusTextAlign(Left/Center/Right). Existing 6-argument call sites stay source- and binary-compatible via explicit secondary constructors. - HUD centering anchors
HudAnchor.ScreenCenterX/ScreenCenterY/ScreenCenter(resolution- independent via the RectTransform anchor system) andHudSpec.DynamicDefaultRect. IFramework.ScreenWidth/ScreenHeight, refreshed once per frame, for resolution-aware layout.ICombatSpeccombat-spec lookup;MeterRowData.SelfAccent.- Buff/skill
NameDesignfallback; leveledskill_level_idsnow resolve to their base skill names.
Changed
Section titled “Changed”- Icon-only overlay buttons are centered and equal-size.
- Hotkey handling, layout-editor input leak, combat-spec resolution, and theme-name sync (in-game hotfix batch, #2).
- Programmatic
SetRectis clamped on-screen so windows can’t be placed unreachable. BarElementfill width now tracksFraction01.
[1.0.1] - 2026-06-17
Section titled “[1.0.1] - 2026-06-17”- Per-frame
GameObject.Findleak that decayed FPS across dungeon re-entries.
[1.0.0] - 2026-06-08
Section titled “[1.0.0] - 2026-06-08”- In-game plugin launcher overlay
- Party roster for raid groups
[0.9.0] - 2026-05-29
Section titled “[0.9.0] - 2026-05-29”- Native HUD element editor (move / hide)
- PlayerHUD demonstrator